Open problems in AI compute verification
127 problem areas that still need work before verifying AI compute is a reliable technology, drawn from 52 documents and grouped by the outside fields most likely to answer them. Most problems show a quote stating the problem, checked against the source by two independent readers; other documents that raise a problem are listed without quotes. Each problem has its own page, listing its sources, its checkable sub-questions (723 in all) and those not yet stated precisely enough to check (283); a further 10 sub-questions without a source are in the data download only.
This is my best understanding as of 2026-10-27. It does not claim that nobody is working on these problems. The list is not complete; problems found since it was drawn up will be added. Corrections are welcome: dan@canaryinstitute.ai.
Problem titles, questions, statements and sub-questions are our paraphrases; only text in quotation marks is the sources' own. A number in a sub-question either is one of the numbers listed below, cites its source, or is marked as set by this survey; a value not yet set is marked “to be agreed”.
A status describes the most advanced public work we found in the documents we hold that state these problems. We have not yet searched the wider literature for solutions, so a Posed or Analyzed problem may have work we have not seen. Based on multi-agent and manual review, we expect roughly 10% of items to change their status upon closer inspection. Corrections are welcome: dan@canaryinstitute.ai.
- Posed a document states the problem, and no public document we hold reports work on it.
- Analyzed a design, protocol, analysis, model or simulation exists, but nothing has been built, run on hardware, or run on real data.
- Started something built, measured or run on real data bears on the question, but does not yet answer it.
- Done a public document answers every part of the question within a stated scope.
- Failed a test, proof or systematic assessment shows that a named family of mechanisms does not work.
At a glance
Click an area to go to its problems; click a problem to see its status and first source, with a link to its own page of sources and sub-questions. Use “Expand all” before searching the page.
Numbers used in the sub-questions
Where a sub-question uses one of these numbers, it takes it from the source given here.
- Run size: 10^24 FLOP, the "Strict Threshold" above which training runs are prohibited in MIRI's draft international agreement (Scher, Abeyta et al. 2025, Article II). Sub-questions about larger runs keep their own numbers.
- Chip line: 16 H100-equivalents: the agreement's "covered chip cluster" line. This survey's island-size study finds that what the line costs an evader depends strongly on how the cost is measured.
- Detection goal: Detect a covert accumulation of 10^24 FLOP with at least 90% probability before it is complete. Like the IAEA's inspection goal, this applies to the regime's checks over a period taken together; a single check or component can be much weaker. It assumes a cheater's gain is at most nine times the sanction (Avenhaus & Canty 1996).
Tamper evidence, seals, enclosures and containment 6 problems
| ID | Status | Question | Sub-qs |
|---|---|---|---|
| P001 | Started | Seals and tamper-evident monitoring of servers and racksWhat detection and false-accept rates do tamper-evident seals, tags and camera/sensor suites achieve on chips, servers and racks under skilled covert attack? | 7 |
Detection rates of tamper-evident seals, tags and camera/sensor suites on chips, servers and racks (not tamper-respondent enclosures). StatusStarted Amodo's own status report of a built tray-level retrofit sensor suite (cameras with OpenCV intrusion detection, PIR, temperature, pressure, IMU, tamper mesh, lid switch, time-of-flight; STM32 TrustZone controller; power-down response) in an RFD still marked 'idea'; no detection or false-alarm rate, no skilled-attacker test (red-team and stress testing are listed as future work), no dataset. The seals clause has no held test: the 289-seal assessment that found all could be defeated cheaply is named by Aarne, Fist & Withers (2024) but not held, so it is not Failed here; 'TamperSec working on this (?)' in the Proofworks projects list is hearsay; Harack et al.'s glitter-and-photograph suggestion is untested. The headline's 'under skilled covert attack' clause has no work. Amodo's v0.1 system includes 8 countermeasures selected from 38 candidates after reviewing 27 attack vectors. … Cameras (4 per tray, OpenCV-based intrusion detection) … Successfully demonstrated detection using STM built-in logic Posed by
| |||
| P002 | Analyzed | Tamper-respondent enclosuresCan a tamper-respondent enclosure for a high-power accelerator pass liquid cooling and full interconnect while erasing its keys on intrusion, at a low false-trigger rate and a hard-to-defeat threshold? | 9 |
Tamper-respondent enclosures for accelerators: red-teams, cooling, zeroisation, false triggers. Terms: FIPS 140 Level 4: the highest tamper-resistance tier in the US/Canada standard for cryptographic hardware, requiring an enclosure that detects and responds to physical intrusion. StatusAnalyzed Design options only, for enclosures around accelerators dissipating up to 1,200 W: active sensing by capacitance, resistance mesh, radio response or distributed sensors; response by key deletion or permanent disablement (capacitor discharge, antifuses, nanothermite); cooling pass-through by a thin boundary under the cold plate, a contained liquid loop with a boundary heat exchanger, or metalised foam for air; enclosure sizes from chip to rack. Nothing built at this scale, no red team, no false-trigger rate; the authors judge that nation-state attackers can likely compromise the best current enclosures and that the per-device cost of circumvention is hard to estimate. The sensing prototypes the section cites are not held. Cross-references, not evidence for this headline: Immler et al. (2019) built and attacked a PCB-scale PUF cover (enclosure PUFs problem); Amodo's tray-level sensor retrofit with power-down response (seals and sensor suites problem) is not an enclosure. Secure enclosures can ideally be made compatible with both air and direct to chip liquid cooling … To permit cooling, the secure enclosure boundary can either be sandwiched between the chip face and cooling pad, or it can contain both the chip and the cooling pad. If it is sandwiched beneath the cooling pad, it has to be very thin and very thermally conductive. Posed by
| |||
| P003 | Started | Enclosure PUFsCan an enclosure-scale hardware fingerprint (PUF) hold a stable key across years of thermal cycling and vibration while resisting probing attacks? | 4 |
Enclosure-scale PUFs: stability, key generation, probing attacks. Terms: PUF (physically unclonable function): a unique, hard-to-copy hardware fingerprint from manufacturing variations; here measured across a whole enclosure to generate a key. StatusStarted Batteryless flexPCB mesh cover deriving a PUF key for a multi-chip PCB embedded system, aiming at FIPS 140-2 Level 3: statistics over 115 covers, physical attacks, and a -20 to +60 C chamber test on one board with three covers plus accelerated ageing of foils. Not accelerator scale or 700–1,200 W, no vibration test, and multi-year key stability not shown (the authors suggest re-enrolling in the field), so the headline's years-of-cycling clause is open; probing resistance is argued by design (keys not reachable without destroying the PUF) rather than tested against microprobing. Kulp et al. (2024) report the demonstration second-hand; the stability question is posed as open by Harack et al. (2025) and the RAND PUF-enclosures RFD. We demonstrate the feasibility of our concept … Practical results including statistics over a set of 115 flexPCB covers, physical attacks, and environmental testing support our design rationale. Posed by
| |||
| P004 | Started | Package- and board-level tamper detectionCan non-destructive inspection detect package- or board-level tampering - circuit edits, re-encapsulation, added components - at a stated detection and false-alarm rate? | 7 |
Non-destructive detection of package- and board-level tampering. Terms: FIB (focused ion beam): a lab tool that can cut and rewire a chip to insert edits. PDN (power distribution network): the network that carries power from the board's voltage regulator to the chip's transistors; a change in its electrical impedance can signal tampering. W D (WD): the Wasserstein distance, the paper's measure of how far a board's measured impedance has moved from its untampered reference. StatusStarted Non-destructive, measured on hardware: an on-FPGA network analyser reads the board's power-delivery-network impedance and flags an added shunt resistor, removed decoupling capacitors, a nearby EM probe and a polished package on commercial FPGA development boards, one board per tamper event, with a global threshold that gave no false alarms on the authors' own measurements. Not a detection and false-alarm rate across devices; not a data-centre GPU board; circuit edits inside the package (FIB) and re-encapsulation are not covered. Related held work: BackMon (Mosavirik & Tajik 2024) detects removal of the cooling stack exposing a flip-chip die backside, one board, no error rates; Safa et al. 2023 use VNA impedance signatures to spot aged chips. The headline is not answered within a stated scope. The magnitude of the WD metric shows that all tamper events performed in this paper can be distinguished if we use the sensor at the right frequency. Moreover, the detection thresholds for different tamper events, and their corresponding frequency bands, can be selected differently. However, we still can define a global threshold at W D = 3 for the tampering experiments presented in this work without causing any false alarms. Posed by
| |||
| P005 | Analyzed | Facility air-gap and enclosure containmentHas an air-gapped facility or pod been red-teamed against electromagnetic, acoustic, optical and personnel channels, with a verified limit on what can leak out? | 9 |
Air-gap and shielded-enclosure containment of facilities and pods. Terms: air-gap: physically isolating a facility from outside networks. dB attenuation: how much a shield cuts a signal, in decibels (every 10 dB is a tenfold reduction). SCIF: a shielded, access-controlled secure room. StatusAnalyzed Desk study only: a literature survey of physical side channels and an example defence architecture for rack or pod enclosures, with an estimated bill of materials below 1% of cluster cost while tolerating single-digit kbit/s of covert bandwidth; the author calls it a two-week sprint, 'far from conclusive'. Nothing built; no facility or pod red-teamed; no verified leakage limit; the RF attenuation of cheaply welded sheet-metal enclosures is an assumption, not a measurement. Single-computer exfiltration demonstrations held in the collection (Fansmitter acoustic exfiltration at 15 bits/min, Deep-TEMPEST HDMI, Synesthesia screen acoustics) are attack measurements at PC scale, not containment tests, and are cross-references because the headline names facility or pod scale. The Proofworks workshop 'seeming consensus' that 1-bit air-gapping is feasible is an opinion. I present an example of a defense architecture towards the end of this post. It leverages a synergy of cheap enclosures, sensors, jamming, unannounced hardware inspections and physical security Posed by
| |||
| P006 | Started | Sealed and maintenance-free operationWhat spare and redundancy fraction and failure rates allow a sealed cluster to run at near-full capacity for years without physical maintenance? | 6 |
Running sealed clusters without maintenance: spares, failure rates. Terms: MTBF (mean time between failures): the average operating time before a component fails, used to size spares for an unattended cluster. N+k redundancy: a node built with k spare units beyond the N it needs to run. StatusStarted Measured failure rates on real field data: 18,688 NVIDIA K20X GPUs on ORNL's Titan (2012–2019), survival analysis, with about 11,000 boards replaced for a resistor-corrosion defect; corroborated by Bode et al. 2021 on Blue Waters (4,228 K20X GPUs, roughly 0.0035 replacements per GPU-year, no wear-out rise). Limits: decade-old GPUs in maintained, air-cooled HPC machines, not a sealed cluster of current accelerators. The spare and redundancy fraction clause has no work: Dean's 2025 hazard-rate note (0.001–0.18 per year) is a desk estimate drawn from these datasets and gives no spares fraction; the flexHEG, Harack and AISF rows are open questions. We write about the last rework cycle and a reliability analysis of over 100,000 years of GPU lifetimes during Titan’s 6-year-long productive period. Using time between failures analysis and statistical survival analysis techniques, we find that GPU reliability is dependent on heat dissipation to an extent that strongly correlates with detailed nuances of the cooling architecture and job scheduling. Posed by
| |||
Chip identity, PUFs, design-to-silicon assurance and supply chain 11 problems
| ID | Status | Question | Sub-qs |
|---|---|---|---|
| P007 | Analyzed | Ruling out secret chip productionCan secret chip production be ruled out, whether at a secret fab or as extra chips from a known one? | 4 |
Whether secret chip production can be ruled out, at a covert fab or as extra chips from a known one, including by bounding a known fab's output from lithography-tool and wafer-start counts, floor space, power use, trackable materials and census reconciliation against shipment records. StatusAnalyzed A desk estimate conditional on the AI 2040 Plan A regime, in which chip production is monitored at a small number of controlled fabs and upstream suppliers' purchases, inventory and cumulative production are declared and cross-checked layer by layer; concealed domestic production is given a 25% feasibility that the authors call an intuitive guess. Shavit (2023) proposes production-line monitoring on the nuclear-fuel model and tracking the lithography supply chain to detect new fabs, without a precision. No fab's output has been bounded by measurement and no covert-fab detection method has been tested. Concretely, we should be able to keep diversion to approximately zero and have very high confidence that total diversion is less than 100k H100e. Raised in: Harack et al. (2025), Verification for International AI Governance (Oxford AIGI); Scher & Thiergart (2024), Mechanisms to Verify International Agreements About AI Development (MIRI); Aarne & Petrie (2025), International Security Applications of Flexible Hardware-Enabled Guarantees (flexHEG Part 3); Shavit (2023), What Does It Take to Catch a Chinchilla? Verifying Rules on Large-Scale Neural Network Training via Compute Monitoring; Koopmanschap & Barten (2026), How to Catch a GPU: A Taxonomy of Verification and Enforcement Mechanisms for International AI Agreements; Wasil, Reed, Miller & Barnett (2024), Verification Methods for International AI Agreements; Proofworks, Living Verification Documents. | |||
| P008 | Analyzed | Chain of custody from fab to packagingCan a custody scheme detect a wafer or die that goes missing or is swapped between fab and packaging, and identify who last held it? | 5 |
Die-level identity and records (ECID, tamper-evident die-ID commitments, tool-signed process records) plus sealed transport, detecting a missing or substituted die between fab and packaging. Terms: ECID (electronic chip ID): a unique serial number burned into each die, readable electronically over the IEEE 1149.1 (JTAG) test interface. Merkle root: a single short fingerprint summarising a whole set of records, so any later change to any record is detectable. StatusAnalyzed A proposed custody principle only: verifiable transport at each step from fab through dicing and packaging, with liability on the last holder; no mechanism, miss rate or test. Cankaya (2025) describes signed ECID handoffs second-hand from Zhang & Guin (2019), which is not held, and leaves 'who last had custody' for a later post; Scher & Thiergart list CISA-style chain-of-custody documentation as a building block. Nothing deployed for an AI-accelerator supply chain. it is important for the wafers (and later the dies) to be securely and verifiably transported at each step. If these chip precursors ever go missing, responsibility for the violation would lie with the most recent holder. Raised in: Scher & Thiergart (2024), Mechanisms to Verify International Agreements About AI Development (MIRI); Shavit (2023), What Does It Take to Catch a Chinchilla? Verifying Rules on Large-Scale Neural Network Training via Compute Monitoring; Ilhan, Withers, Gietz & Harack (2026), Verifiable Semiconductor Manufacturing (Oxford AIGI); Cankaya (2025), TSMC Most Definitely Has a Golden Record of All AI Chips It Made. | |||
| P009 | Analyzed | Production oversight of chips and mechanism componentsWhat institutional oversight - monitoring, inspection standards, independent testing and verified destruction - covers the production of chips and of the components that verification mechanisms depend on? | 8 |
Institutional oversight of chip and mechanism-component production: monitoring regimes, inspection standards and capacity, independent testing, verified destruction of failed units, tool-use controls, and retroactively establishing trust in already-produced chips. Terms: p: the fraction of devices sampled for imaging. d: the chance that imaging catches a compromised device it inspects. n: the number of compromised devices. StatusAnalyzed A design for an international body's oversight of guarantee-processor production: manufacturing equipment locked to the approved design, independent testing by a party other than the manufacturer, continuous camera and on-site staff oversight, tamper-evident shipping, extensive testing and imaging of a 1/1000 sample (the authors' own arithmetic gives a 63% chance of catching a 1,000-device compromised batch), verifiable destruction of failed devices, published processes and authorised penetration tests. Nothing instituted or piloted. Oversight of accelerator fabs themselves is only an analogy to nuclear-fuel monitoring (Shavit 2023); retroactive trust in already-produced chips is a proposal to sample and destructively reverse-engineer existing components (flexHEG Part 1). No held document reports any oversight regime in operation. Each guarantee processor should be subjected to automated, independent testing. … the testing should be done by a different party than the one that manufactured it, ideally either by the international body, several member states, or both. Posed by
| |||
| P010 | Started | Chip identity and fingerprintingIs there a clone-resistant chip or board identifier, or a fingerprint that tells apart individual chips of the same model, with a published false-match rate and cost that a prover cannot defeat by swapping in another chip? | 7 |
Clone-resistant chip/board identifiers and same-SKU die fingerprints. Terms: same-SKU die: an individual chip of the same product model as others (SKU = a product/stock code). PUF (physically unclonable function): a chip's unique, hard-to-copy fingerprint arising from tiny manufacturing variations. S-parameters: radio-frequency measurements of a board, usable as such a fingerprint. StatusStarted Software (atomic-race) fingerprint telling apart 24 same-model H200 GPUs rented on vast.ai, 480 runs; up to 100% paired re-identification. The authors' own limitations: not validated at larger scale; stability over time, transport and temperature not established; security against an adversary who emulates the GPU or runs the function faster not tested, so the headline's 'cannot defeat by swapping in another chip' clause and a published false-match rate at scale remain open. Lower-ranked evidence: Alpay & Alpay (2026) measure a per-SM latency map stable over a six-hour RTX 5090 run and 100% leave-one-out separation across distinct Blackwell dies, but leave same-model die separation to prior work; Safa, Mosavirik & Tajik (2023) distinguish an aged STM32 from 10–12 genuine ones by S-parameters with no error rates. No cost figure in any held source. we evaluated the function across 24 Nvidia H200 GPUs, 2 seeds, and 10 runs each … fingerprinting function that achieved up to 100% accuracy in small-scale tests Posed by
| |||
| P011 | Analyzed | Matching fabricated silicon to the approved designCan a party other than designer and fab confirm fabricated silicon matches the approved design and detect a deliberately inserted mismatch at a stated rate? | 8 |
Checking fabricated silicon against the approved design. StatusAnalyzed A proposal only: the non-manufacturing side samples fabricated components and destructively reverse-engineers them against the manufacturer's design documents; flexHEG Part 2 adds open-sourcing the guarantee-processor design for comparison with physical scans of randomly selected chips, and Ilhan et al. (2026) analyse physical inspection and sampling as the direct check. No held document reports a third party checking a fabricated die against its design, and none gives a detection rate on a deliberately inserted mismatch; the only rate is Agrawal et al. (2007), a circuit-level power simulation (49/49 simulated Trojan chips caught, 1/49 false alarm) on a 28k-gate RSA core, no silicon measured. Puschner et al.'s real-world SEM Trojan-detection case study is named by the Proofworks neutral-compute note but not held. some of the existing components could be sampled for rigorous, destructive reverse-engineering to check that they match the design documents provided by the manufacturer. Posed by
| |||
| P012 | Analyzed | Physical inspection capability and costWhat is the per-die cost, time and miss rate of destructive and non-destructive die inspection, and how many chips must be inspected to catch an inserted trojan with high confidence? | 8 |
Capability, cost and sampling statistics of physical die inspection. Terms: q: the fraction of chips that carry a hardware trojan. StatusAnalyzed A desk analysis of the sampling clause: inspection cost and sample size against Trojan prevalence, with the observation that sampling 1% of chips may cost more than producing all chips unverified; Aarne & Petrie (2025, flexHEG Part 3) give the same arithmetic for a 1/1000 sampling rate (63% chance of catching a 1,000-device compromised batch). No held document measures per-die cost, time or miss rate of destructive or non-destructive inspection at an advanced node: Harack et al. (2025) state no non-destructive technique is known for a chip with tens of billions of transistors, and Ilhan et al. call inspection 'extremely expensive, slow, and often destructive' without numbers. Cross-references, not evidence for this headline: Agrawal et al. (2007) simulate side-channel screening (no silicon); Thummala et al. (2026) measure a power side-channel screen on one Atmel microcontroller (93.2% of tampered-firmware traces flagged at 1% false positives), firmware tampering rather than die modification. If only 1 chip in 100,000 is compromised, many thousands must be tested to achieve high confidence of detection. The sampling regime can be tuned based on acceptable risk (e.g., testing 1% of chips vs 0.01%), but this involves significant cost considerations. Posed by
| |||
| P013 | Analyzed | Designer backdoors and hidden functionalityCan a verifier gain assurance that a chip design is free of designer-inserted backdoors or hidden execution paths, at a stated detection rate on seeded trojans? | 3 |
Detecting designer backdoors and hidden functionality. StatusAnalyzed Design options only, for the guarantee processor: open-sourcing the design and logic for third-party analysis against physical scans of randomly selected chips, executing guarantee logic on publicly auditable FPGAs, and running redundant guarantee processors from different manufacturers in lock-step; flexHEG Part 3 adds collaborative design with justification of every choice. Nothing built, no outside review of real silicon, and no detection rate on seeded trojans anywhere in the held sources; the binning clause (dormant functionality) has only RAND's design rule that there be none. To increase trust that backdoors have not been added to the Guarantee Processor design or manufacturing process, some options are: … Open-sourcing the guarantee processor design and guarantee logic, which could then be analyzed and compared with physical scans of randomly selected chips. Posed by
| |||
| P014 | Analyzed | Design-flow and fab-process integrityWhat detection rate do automated trojan detectors reach on design-flow and fab-process data for unseen trojan families at accelerator scale? | 4 |
Trojan detection in design flow and fab process data. StatusAnalyzed A desk analysis of the fab-process clause: repurposing fabs' existing process-data anomaly detection and virtual metrology for security, with its stated limits (catches only attacks that show as statistical deviations; high false-positive risk in a noisy fab). No detection rate anywhere. The design-flow clause rests on graph-neural-network detectors trained on netlists reported second-hand (Thorat et al. 2025, not held), which flag patterns resembling known Trojans rather than unseen families; no held document tests any detector at accelerator scale or on unseen trojan families. Logic locking and other obfuscation are described in the same paper as broken by SAT-style attacks, but those attack papers are not held. A more scalable approach is to integrate verification into the fabrication process itself. Modern fabs already collect enormous amounts of process data for yield and quality control, including sensor readings, optical inspections of masks and wafers, and electrical test parameters. By applying statistical anomaly detection to this data, one can attempt to catch deviations that might indicate tampering. Posed by
| |||
| P015 | Posed | Advanced packaging (chiplets, 3D stacking)Can covert channels, untrusted-die key access, or die substitution be detected in chiplet / 3D-stacked packages at a demonstrated rate? | 3 |
Security of chiplets and 3D-stacked packages. Terms: chiplet: a chip built from several smaller dies packaged together. 3D stacking: stacking dies vertically in one package. UCIe: an industry standard for linking chiplets inside a package. StatusPosed Posed by
| |||
| P016 | Analyzed | Supply-chain assurance overall and neutral hardwareCan an end-to-end chain of assurances cover design, fabrication, packaging and deployment, and who can credibly produce neutral verification hardware? | 0 |
End-to-end supply-chain assurance and who can build neutral hardware. StatusAnalyzed A design answer to the neutral-producer clause: redundant guarantee processors from each bloc and nested enclosures each built by a different bloc, with the integration details left out of scope; the Proofworks neutral-compute note sketches the alternative of open-sourcing designs down to GDSII and fabricating on trailing nodes, which it judges impossible at the leading edge. The end-to-end chain clause has Ilhan et al.'s (2026) analysis of layered design-phase and manufacturing-phase measures, with the chain itself posed as an open question. Nothing built; no neutral producer identified (the Proofworks prospects note sees no candidate in a US–China scenario). One colour: both clauses are answered by the same hardware supply-chain security community, so the sorting-pass split is not used. they might instead rely on a design where different countries provide different overlapping components of the design, such that each country can trust the system without trusting all of the components. … security processors could be redundant, such that everything passes through two chips that implement the same specifications, one designed and manufactured by each bloc. Posed by
| |||
| P017 | Analyzed | Legacy installed base and retrofitWhich verification functions can be retrofitted onto the existing installed base of accelerators by firmware alone or an external module, and what fraction of installed compute cannot take them? | 1 |
Retrofitting verification functions onto the legacy installed base. Terms: H100 / A100: Nvidia's current and prior-generation data-centre GPUs, used here to stand for the large existing installed base of accelerators. StatusAnalyzed A desk analysis drawing on industry conversations: which governance mechanisms the installed base could take by firmware alone (the H100's security module, TEE and secure boot are noted as present), with timeline estimates by threat context (months for firmware-only, 18 months to 4 years for silicon changes, several years for tamper-proofing) and the caveat that existing secure-boot implementations likely carry unpatchable flaws. External-module retrofit is a design option in flexHEG Part 2 (capacitor-based disablement 'a good option for retrofitting'; antifuses unsuitable). No held document demonstrates a retrofit or states what fraction of installed compute cannot take one; Shavit (2023) notes many sold chips may be neither retrofittable nor locatable. Brass & Aarne (2024), cited by Tee & Happel for delay-based location verification on existing confidential-computing keys, is not held. For minimally adversarial contexts the current level of hardware security likely would be sufficient, and thus many mechanisms could be implemented as firmware updates. This would take a few months. Some mechanisms may not be possible to implement this way on current hardware, in which case silicon-level changes would be required, and the time to implement them would increase to between 18 months and 4 years. Posed by
| |||
Side channels, on-device power/EM measurement and telemetry integrity 5 problems
| ID | Status | Question | Sub-qs |
|---|---|---|---|
| P018 | Analyzed | Telemetry read-path integrity and accuracyDoes any shipping GPU expose an authenticated, tamper-resistant power/clock telemetry path a root-privileged host cannot forge or replay? | 6 |
Authenticity and accuracy of GPU power/clock telemetry read paths. Terms: NVML (Nvidia Management Library): the standard software interface for reading a GPU's power, clock and temperature. StatusAnalyzed A requirements design for the read path (attested enclave or secure-boot-chained firmware, hardware-rooted signing, nonce-binding against replay, attested monitoring software), pointing to the flexHEG guarantee-processor designs; nothing built, and no shipping GPU is named as having it - every source says such paths 'do not yet exist'. Telemetry accuracy (sampling fraction, counter error) is in the statement but not the headline: related, with measured work (Yang et al. 2023: only 25% of runtime sampled for power on A100/H100). The monitoring agent must read counter values from a source the adversary cannot tamper with—e.g., an attested enclave on the GPU firmware boundary (Aarne et al., 2025) or a verified read path through secure-boot-chained firmware (Petrie, 2025; O’Gara et al., 2025). Without this, an adversary with software privilege simply returns forged values. Posed by
| |||
| P019 | Started | Bounding compute from power and analog measurementHow tightly can power and analog observables bound a chip's or facility's compute, and how much does the bound loosen from on-die to feed meter under adversarial choice? | 23 |
Power-side compute bounds (beta*, energy per op) and the chip/node analog observables and meter limits that feed them, from on-die to feed meter. Terms: β (beta): the largest hidden computation a power trace cannot rule out, as a fraction of the declared machine capacity. It depends on declared utilisation (how busy the declared work keeps the chip); β (beta) is its worst case over all utilisations. The paper's headline "β = 1.16" is that worst case; 0.41 is what a demonstrated attack actually hid. Energy per operation: joules consumed per arithmetic operation, used to turn power readings into a compute estimate. ΔP0 (delta-P0): the spread of the chip's non-compute overhead power (idle draw, cooling, I/O), which the verifier cannot see separately from compute; 41.2 W on the A100 measured. w_C: how much energy per operation varies with where the data sits (the 'locality' factor; 1.04× on the A100 measured). N: the number of devices. StatusStarted One NVIDIA A100 model, one driver, one cluster; energy read from on-die telemetry with a perfect meter assumed; the authors' own adversary model; the bound tightens to about 0.06 only with re-execution at an observed operating point; a 0.41-1.16 gap left open. The loosening from on-die to feed meter and the multi-device facility bound are explicitly not addressed ('assume a perfect meter'; 'do not offer a facility-level, multi-device version'), so that clause has no work. Measurements on NVIDIA A100 GPUs constrain β = 1.16 in the worst case, while adversarial matched-energy strategies are shown to hide at least β = 0.41 of compute. Posed by
| |||
| P020 | Analyzed | Side-channel defences at scaleCan side-channel defences (shielding, filters, jamming, wardens) be mass-manufactured and retrofitted to cap a tray's leakage below a stated level, and at what unit cost? | 7 |
Engineering side-channel defences (TEMPEST, filters, jamming, wardens) at scale. Terms: TEMPEST: the standard body of techniques for shielding equipment so it does not leak information through stray electromagnetic emissions. wardens: monitors that watch for such leakage. StatusAnalyzed A paper design from a research sprint the author calls 'a first-pass viability study': enclosures, sensors, look-through jamming, inspections and physical security with an estimated bill of materials below 1% of cluster cost (this survey's cost goal of 1% of facility cost, met on paper only); power-filter volume cost 'has proven difficult' to pin down; no attenuation or leakage-rate cap measured; the 39 dB look-through suppression is cited from commodity radios in work not held; wardens: no work ('unexplored'). I present an example of a defense architecture towards the end of this post. It leverages a synergy of cheap enclosures, sensors, jamming, unannounced hardware inspections and physical security … at an estimated (!) Bill-of-Materials (BOM) below 1% of the cost of the monitored cluster Posed by
| |||
| P021 | Started | Trustworthiness and leakage of monitoring sensors and dataCan facility sensors be built that both parties trust — verifier trusts the readings, prover trusts they do not exfiltrate — with a quantified leak bound? | 5 |
Whether facility sensors and their data can be trusted and leak little. StatusStarted Leakage of verifier-facing releases derived from real physical measurements (H100 and B200 node power traces, RAPL power traces, GPU electromagnetic traces, measured workload profiles) over six verification tasks, measured as conditional mutual information: this covers the prover's quantified leak bound for what the verifier receives. The verifier-trusts-the-readings clause is out of scope ('MID does not make a compromised evidence source honest'); exfiltration by the sensor hardware itself (bits per second per channel) and a mutually trusted sensor: nothing built; Harack's prover-side mitigations are paper only. Across these tasks, MID produces three releases with perfect held-out verification and zero measured collateral leakage, while the remaining tasks yield explicit privacy–utility frontiers. Posed by
| |||
| P022 | Started | Training-run power signatureWhat are the iteration period, dip duty cycle and harmonic structure of a frontier training run's power trace at chip, rack and facility meter? | 4 |
The power signature of a training run at chip, rack and facility meter. StatusStarted One normalized production training trace on DGX-H100 racks (no GPU count, model or absolute power): a 0.2-3 Hz band and an iteration dip recurring 'from once per second or less, to once every tens of seconds'. Dip duty cycle, step-time jitter and harmonic amplitudes at chip, rack and facility meter are not reported. Node level: Vercellino et al. (2026), also held, measure 4xH100 node power at 0.1 s for fine-tuning and training up to 16 nodes. Facility meter: simulation only (Ko & Zhu 2026, whose assumed 0.5-1.5 Hz iteration rate conflicts with Choukse; Vercellino's stitched facility profiles) and an unattributed once-a-second grid sag at a data-center substation (Mishra et al. 2025). AI workload power traces like those in Figure 1 show FFT energy concentrated between 0.2–3 Hz (Figure 3)—close to known resonant modes of turbine-generator shafts and long transmission lines. Posed by
| |||
Covert channels, steganography, egress and sealed-environment data flow 3 problems
| ID | Status | Question | Sub-qs |
|---|---|---|---|
| P023 | Started | Residual covert-channel capacity out of monitored computeHow many bits per second or per token survive monitoring via steganography, protocol free variables, and physical channels (fibre, coolant, RF)? | 18 |
Measured bits/s or bits/token that survive monitoring: steganography in outputs and egress, protocol free variables, and physical channels (fibre, coolant, RF). StatusStarted Steganography in inference outputs only: the authors' own detector on open-weight models of 3B-30B parameters under their trust assumptions (every input, output and seed logged; honest verification server). Protocol free variables: Cankaya's Fermi estimate of about 100 bits per second through active north-south taps, which he calls informed guesses needing validation - paper only. Physical channels (fibre, coolant, RF): Cankaya's side-channel survey estimates ('we expect Mbps to be possible'; 'strong case studies are few') - paper only; NVLink contention: no work. Older covert-channel measurements in the collection (Ethernet padding, thermal) were not made against compute monitoring. On MOE-Qwen-30B, our detector reduces exfiltratable information to < 0.5% with false-positive rate of < 10−2 , corresponding to a > 200× slowdown for adversaries. Posed by
| |||
| P024 | Analyzed | Data in and out of sealed verification environmentsCan data be moved into a sealed verification facility, and results out, while bounding what leaks, including through faults and retries? | 5 |
Protocols for moving data into, deleting it from, and bounding leakage out of a sealed auditing facility (including one-bit outputs and fault budgets). Terms: SCIF (sensitive compartmented information facility): a shielded, access-controlled room built to stop information leaking out. air-gap: physically isolating a system from outside networks. StatusAnalyzed Design only: a neutrally located, mutually monitored and possibly air-gapped data centre runs privacy-preserving evaluations, assessors emit one bit per judgment, and the sensitive data is verifiably deleted afterwards (Fig. 2.1 caption); how the bit is transferred is left to future work (fn 261). The fault-and-retry leakage budget appears only as a stated requirement in Cankaya's self-deprecated System Overview v0.2 (a pre-agreed budgeted fault rate). Nothing built, no red-team, no measured leak bound. After conducting their evaluations as agreed by the Prover and Verifier, the assessor would be able to reliably transmit one bit per judgment to the Verifier and Prover. Raised in: Harack et al. (2025), Verification for International AI Governance (Oxford AIGI); Cankaya (2026), A System Overview for Near-Term, Low-Trust AI Compute Verification (v0.2); Cankaya (2026), All Hands on Deck to Build the Datacenter Lie Detector; Cankaya (2026), Research Note: The Fundamentals and Feasibility of Secure Network Taps for Verifying AI Datacenter Use. | |||
| P025 | Started | Digital perimeters, egress accounting and inference-site architectureCan a digital-perimeter and egress-accounting design show, to stated confidence, that every outbound message from an inference site has a log entry? | 8 |
Digital perimeters, egress accounting and inference-site architecture. StatusStarted Hashing every frame on one 400GbE link with generator traffic on real hardware: line rate holds for frames of 1500 B and above on a 32-core CPU, fails for 64 B frames and on DPU offload - a prover choosing small frames defeats complete hashing. This is the capture layer only; reconciling hashed egress against an inference log to a stated confidence, red-team exfiltration tests, and the inference-site architecture (weights path over partitioned networks and data diodes, where 'we have not seen any work') are paper only or no work. The designs are Cankaya's north-south taps with all egress 'fully explainable' by observed activity and Rinberg's Merkle inference ledger. The same benchmark is the evidence for the traffic-hashing problem. For worst-case 64B minimum frames, no algorithm on this hardware reaches full line rate … For realistic traffic profiles at 1500B and above, AES-128-GMAC and SipHash-1-3-128 are both capable of sustaining line-rate hashing Posed by
| |||
Optical taps, line-rate capture, network measurement and interconnect attestation 12 problems
| ID | Status | Question | Sub-qs |
|---|---|---|---|
| P026 | Started | Location verification of chipsWhat location precision and false-accept rate can network-delay geolocation of accelerators achieve against an adversary who adds delay, relays, or spoofs? | 6 |
Delay-based and network-measurement geolocation and co-location of accelerators against adversarial delay, relays and spoofing. Terms: N: the number of chips taking part in a co-location challenge. StatusStarted The authors' own measurement of their own method on rented cloud GPUs with cooperative hosts: network-delay geolocation against public RIPE Atlas landmarks places a B200 within 44 km and a V100 within 270 km (79-298 km depending on landmark density) and rejects 11 of 11 decoy sites; the decoys are other locations, not an adversary adding delay or relaying, and the paper says a high round-trip time is no evidence of absence from a forbidden region and the scheme is not tamper-proof, so the headline's false-accept rate against added delay, relays or spoofing is not reported. Timed co-location challenges across N chips: no work. Robustness of on-chip GPS to commercial spoofers is a related sub-question, not in the headline, with no work. 169 RIPE Atlas probes place the B200 within 44 km of its claimed datacentre and reject all 11 decoy sites. Posed by
| |||
| P027 | Started | Enforcing and attesting cluster composition and interconnect limitsCan hardware fix which devices may link, cap inter-device bandwidth, and attest cluster size against bridging devices, at what training slowdown? | 14 |
Hardware that fixes which devices may link, caps inter-device bandwidth, and attests cluster size/configuration against bridging devices. StatusStarted Amodo's own report of its own work, not independently checked: a central controller leases bandwidth and transfer-volume quotas to BlueField DPUs over an encrypted tunnel, the DPUs enforce them in software with the Linux traffic controller, and the system runs on one research server; hardware acceleration on BlueField-2/3 is listed as in progress. Not shown: a hardware-fixed set of permitted peers, attestation of cluster size against a bridging device or reconfiguration, holding a cap at line rate (the RFD lists 'can software tc keep up?' as open), and any training slowdown. Fixing which devices may link and attesting cluster size: paper designs only (RAND 2024 proposes restricting interconnect to a handful of preauthorized chips and says the mechanism 'does not yet exist'; flexHEG proposes verifying cluster configuration and outgoing bandwidth). Training slowdown when traffic is forced off NVLink: no work (RAND's 'might not present a significant performance penalty' is speculation). "Phase 1: Software implementation (complete)" … "DPUs enforce quotas via Linux traffic controller (tc)" … "Running on Amodo research server" Posed by
| |||
| P028 | Analyzed | Taps on scale-up and electrical interconnects (NVLink)Can traffic on high-speed electrical links between accelerators (such as NVLink/NVSwitch) and other non-optical interconnects be captured or witnessed at full line rate by a trustworthy add-on device? | 6 |
Capturing or witnessing NVLink/NVSwitch, copper and other non-optical interconnect traffic at line rate. Terms: NVLink / NVSwitch: Nvidia's high-speed electrical links and switch chips that connect GPUs within a server or rack. StatusAnalyzed Desk analysis of two designs for copper (NVLink-class) links, a modified switch ASIC that copies sampled traffic from its digital fabric and a bump-in-wire retimer exposing its equalized bitstream, with the author's expectations of about 60 ns added per hop, dependence on the switch or retimer vendor, and at least a year of lead time; nothing built or measured. Milton & Reynolds (Amodo, 2026) likewise describe an optical-electrical-optical retimer tap and an electrical interposer at the transceiver port as designs 'not commercially available' that need to be designed and tested. Full line rate and the trustworthiness of an active device: argued (active electronics add attack surface), not tested. Bidirectional circulator-optics TPU pod links: no work. Peigne-Lefebvre et al. (2026) state NVLink traffic has no physical witness. Bump-in-wire retimer tap: Insert a signal retimer into the copper link and extract a digital copy from its internal equalized bitstream. Posed by
| |||
| P029 | Analyzed | Passive optical taps at high lane ratesCan a passive optical tap and monitor receiver recover a high-speed optical lane within its link budget while keeping the live link error-free? | 7 |
Passive optical taps and monitor receivers at 53-112 GBaud. Terms: GBaud (gigabaud): billions of signal symbols per second. PAM4: a scheme sending two bits per symbol using four levels. FEC (forward error correction): extra data that lets a receiver fix errors. DR4/DR8: standard multi-lane optical link types. StatusAnalyzed Desk link-budget analysis for the live leg: baseline optical margins against tap insertion loss, with feasibility thresholds of 1.5 dB and 0.5 dB, for 400G-class multimode links inside a scalable unit and single-mode links between units; the monitor-leg recovery at 53-112 GBaud is left as the harder problem and the author says amplification is probably indispensable at 112 GBaud. No tap built or measured at this problem's lane rates. Milton & Reynolds (Amodo, 2026) reach a similar desk result ('even a 1 dB loss may be unacceptable') and say they plan testing; commercial passive taps exist for 26 GBaud lanes (400G-DR8), below this problem's lane rates, which are the headline's stated difficulty, so that is a cross-reference and not evidence; vendors' statements that 53 GBaud should work are unverified hearsay. Cankaya's fleet-margin argument cites Meta production telemetry (John et al., DTS 2025), not held. Intra-SU multi-mode: Passive taps feasible only at 95/5 with direct cabling. Infeasible with standard structured cabling. Posed by
| |||
| P030 | Analyzed | Tap placement and coverageWhere must taps sit to cover the communication paths that matter, and what fraction of compute can run undeclared work invisible to taps at a given network level? | 8 |
Where taps must sit and what coverage they give. Terms: q: the probability that an evader's traffic escapes the taps. StatusAnalyzed Desk analysis of tap placement (north-south edge, storage fabric, east-west levels) and what each level misses; names copper scale-up racks (NVL72, NVL576) and TPU pods above 9,000 chips as blind spots where a frontier run could complete without crossing a tap; the covert-bandwidth figures are Fermi estimates the author calls informed guesses; no fraction-of-compute figure and no measurement; the author lists a rigorous threat model of covert workloads under each placement as his most important open question. For most deployments, we expect spine-to-leaf to be the lowest practical level for tapping intra-SU/pod compute fabric. Posed by
| |||
| P031 | Started | Line-rate capture and hashing hardwareCan hardware hash, timestamp and sign every packet at full line rate with zero loss, and at what resource and power cost? | 7 |
Line-rate capture, hashing and signing hardware. Terms: FPGA (field-programmable gate array): a chip whose logic can be reconfigured, often used for line-rate packet processing. DPU / SmartNIC: a network card with its own processor that can inspect and process traffic. StatusStarted The authors' own CPU/DPDK benchmark at 400 Gb/s: line-rate hashing for frames of 1500 B and above; at 64 B minimum frames no algorithm reaches line rate within the physical core count; hashing only, with no timestamping or signing, zero loss on real datacenter traffic not shown, and no trusted or simple hardware; not 800G or 1.6T (limited hardware availability); the authors say an FPGA or ASIC is likely needed. Peigne-Lefebvre et al. (2026) still pose formally verified FPGA hashing at 400+ Gbps as open. For realistic traffic profiles at 1500B and above, AES-128-GMAC and SipHash-1-3-128 are both capable of sustaining line-rate hashing Posed by
| |||
| P032 | Analyzed | Tap trust under mutual distrustCan a tap and its trusted core run at line rate while staying simple enough that two mutually distrusting parties can inspect and trust it, with no room for a hidden backdoor? | 7 |
Trusting tap hardware under mutual distrust. Terms: TCB (trusted computing base): the minimal set of hardware and software that must be trusted for a security claim to hold. data diode: a link that physically allows data to flow in only one direction. StatusAnalyzed A protocol design: tap logic kept minimal and physically inspectable, with the prover tapping the tap's own egress so that processors trusted by one side only can suffice ('watching the watchers'; the note points to a separate design built from commodity hardware with redundant processors). The same note says single-party active wardens 'have been built and tested' in prior work (not held, not named) and that extending them to mutual distrust is non-trivial. Nothing built; no line rate shown; formal verification of tap logic, TLS termination in software-free logic, and a secure alarm channel between distrusting processors are left open; Cankaya's System Overview v0.2 says an FPGA active warden has not been made yet. passive optical splitters on tap egress links can let the Prover observe every bit sent to the Verifier, enabling independent re-computation to confirm taps behave as specified. Posed by
| |||
| P033 | Started | Reproducibility and canonical definition of tapped trafficCan tapped traffic be defined precisely enough (hash boundaries, payload depth) that both parties reproduce identical hashes from replay plus declared state? | 5 |
Defining tapped traffic so both parties can reproduce its hashes. StatusStarted The author's own early simulations of network-traffic reproducibility, reported and not published, excluding ML non-determinism; no hash-boundary or payload definition has been specified (the note calls an implementable protocol 'a natural next step'); Amodo's SITREP tags network reproducibility 'Not started' and Reynolds & Foster call the payload definition the most important open question. Can network traffic (excluding ML non-determinism) be reliably reproduced? Early simulation work looks promising but needs further validation. Posed by
| |||
| P034 | Analyzed | Attributing captured traffic to declared workloadsCan analysis of captured traffic attribute it to declared workloads and detect an undeclared workload embedded in a declared one, at stated error rates? | 3 |
Attributing captured traffic to declared workloads. StatusAnalyzed A protocol design with a security argument: an auditor-owned passive tap hashes inter-node collective payloads into a Merkle tree that is checked against the declared training at challenge time; the authors list the canonical wire-to-tensor mapping ('not a stable public interface today'), auditing that mapping without trusting the trainer's host, and the absence of any NVLink witness as open problems; the paper's prototype concerns the proof system, not the tap or the attribution. Cankaya (2026) separately proposes per-transfer metadata headers (job ID, tensor type, layer, step) so captured samples can be matched to declared workloads, with negligible overhead asserted, not measured. Detecting an undeclared workload embedded in a declared one at stated error rates: no measurement on real traffic; Baker et al. (2025) call the analysis an unsolved problem. Network anchor. Catches any deviation that changes wire-bound content; deterministic (not statistical) detection on any sampled wire-bound tensor. Posed by
| |||
| P035 | Analyzed | Capture volume, storage and recomputation infrastructureWhat capture bandwidth, storage and recomputation infrastructure does tap-based verification require, and can sampling cut it to a feasible level? | 6 |
Capture volume, storage and recomputation infrastructure. StatusAnalyzed Desk estimates of the quantities asked: sampling 0.01% of a scalable unit's 256 400G scale-out links through a 1x256 optical circuit switch to one FPGA, about $25,000 per unit or about 0.2% of an inferred $12M unit price; storage-fabric capture sized from bursty traffic (about 40 GB/s peak, 3 GB/s average); Amodo's SITREP gives a worst case of one recomputation NIC per inference front-end NIC and tags sampling taps 'Not started'; the author reports that typical north-south edge traffic composition was unavailable to him. Nothing built or measured; this survey's cost goal of 1% of facility cost is met only on paper. A sampling rate of 0.01% of the compute fabric traffic would mean about 20Gb/s of traffic to timestamp and hash (and maybe parse). Posed by
| |||
| P036 | Analyzed | Tap supply chain and manufacturing scaleCan taps and their components (splitters, optical switches, transceivers, programmable-logic chips) be manufactured at the scale a global tapping regime needs, and how quickly? | 5 |
Manufacturing scale of taps and their components. Terms: OCS (optical circuit switch): a switch that routes light paths directly. MEMS: the micro-mechanical mirrors used inside some optical switches. SMF (single-mode fibre): the standard long-distance optical fibre. StatusAnalyzed A Fermi estimate of global tap-component demand against market-forecast production: optical circuit switches about 10K units in 2023 and over 50K projected for 2029, passive splitters in the hundreds of millions, PAM4 DSPs tens of millions, FPGA volumes 'uncertain'; a phased multi-year rollout, not a dated schedule. The author adds that OCS supply appears demand-constrained and can scale with demand. Nothing manufactured or measured. Conclusion: The most critical bottlenecks are in the supply of optical circuit switches and SMF transceivers. Both are already struggling to serve existing demand. Posed by
| |||
| P037 | Analyzed | Encrypted traffic and tapsCan a tap verify encrypted inter-accelerator traffic through confined key disclosure, and at what security trade-off and throughput cost? | 4 |
Taps versus encrypted interconnect traffic. Terms: encrypted interconnect traffic (TEE/CC): data sent between accelerators encrypted by a trusted execution environment or confidential-computing feature. confined key disclosure: giving a tap just enough key material to check traffic without exposing everything. StatusAnalyzed Desk analysis of key disclosure for tapped encrypted traffic: where encryption runs outside a trusted execution environment the prover can retain session keys and share them after sessions expire; for TEE-encrypted fabrics post-hoc decryption is called fundamentally difficult and an active man-in-the-middle tap is sketched as a possible workaround with unestimated overhead; the security trade-off is argued (taps as a substitute for TEE isolation against a malicious operator), not tested. Throughput cost of any such tap: no work. Security Level 5 encrypted-traffic cruxes were found at a workshop and not resolved. This means that for non-TEE encrypted fabrics, the operator (i.e., the prover) can retain keys to enable post-hoc decryption of tapped ciphertext. Encryption alone does not foreclose verification. Posed by
| |||
Grid and facility sensing, remote sensing and undeclared-facility detection 11 problems
| ID | Status | Question | Sub-qs |
|---|---|---|---|
| P038 | Analyzed | Facility power modelsHow large is the error in facility- and distribution-level power models when the operator chooses workloads to mislead them? | 7 |
Validity of facility- and PDU-level power models under adversarial choice. Terms: PDU (power distribution unit): the rack-level equipment that feeds and meters power to servers; a facility/PDU power model estimates compute from power readings at those points. StatusAnalyzed Honest operator; facility power modelled, not metered (the NLR dataset combines measured workload profiles with facility power from a bottom-up model of the power-delivery components); 2-16 Kestrel nodes of four H100s each running Llama-2 70B LoRA and Stable Diffusion: a node-count report from mean facility power tolerates about plus or minus 15% uniform calibration error. Adversarial workload choice, the headline's condition, is not modelled at facility or distribution level. Cross-references, not evidence, because the headline names facility scale: Kimpson, Baker & Graham's adversarial power bound is for a single A100 and the paper says it offers no facility-level, multi-device version; Rahman & Tajdari measured wall versus NVML power on one workstation under evasion strategies. When every power value is multiplied by the same calibration factor, the fixed quantizer remains perfect from −15.5% to +18.5% error; failures begin at −16% and +19%. This sweep measures tolerance to uniform calibration error. Posed by
| |||
| P039 | Started | Power accounting at declared facilitiesWhat error does power accounting give for the count of operating AI chips or the size of a run at a declared facility, under adversarial choice? | 4 |
Power accounting of chips and runs at declared facilities. Terms: β (beta): the largest hidden computation a power trace cannot rule out, as a fraction of the declared machine capacity; β (beta) is its worst case over declared utilisation (how busy the declared work keeps the chip). The paper's headline 1.16 is that worst case on the A100s measured; 0.41 is what a demonstrated attack actually hid. StatusStarted Measured on one NVIDIA A100 (energy read from the on-die counter, not a feed meter): a power trace alone cannot rule out between 0.41 and 1.16 machines of undeclared compute per declared machine, against an adversary who chooses numeric format, operand values and idle state; same-SKU transfer checked on three more A100s. One device only, so the facility-level, mixed-fleet and run-size clauses of the headline have no measured bound; time-resolved signals (training-step periodicity) are out of the paper's scope. Lower stages: Baker et al.'s PDU metering and three compute-accounting options (RAND 2025, App. A.6) are a design only; the synchronous-swing and shutdown-verification clauses are stated judgements with no work. A companion simulation study by the same group (October 2026) tracks the training-step rhythm in facility power instead of total energy; it detects synchronous training far more reliably but does not say how much compute is running. "Measurements on NVIDIA A100 GPUs constrain β = 1.16 in the worst case, while adversarial matched-energy strategies are shown to hide at least β = 0.41 of compute." Posed by
| |||
| P040 | Analyzed | Grid-side detection of small hidden loadsWhat is the smallest undeclared AI load detectable from grid or utility interval data at a fixed false-alarm rate? | 10 |
Detecting small hidden AI loads from grid and utility data. Terms: feeder: a distribution power line serving an area. phase unbalance: unevenness across the three phases of AC power, a tell-tale of some loads. CUSUM: a statistical test that flags a sustained shift in a signal. StatusAnalyzed A desk reference-class analysis of utility electricity records by facility size, using Pilz & Heim's data-centre counts and EIA survey counts: a 10 MW continuous load is visible at the utility level but blends with crypto-mining, continuous-process industrial and campus loads; at 1 MW and 100 kW the paper places the load among flat-load commercial customers "of which the United States has millions", a figure it takes from a count of commercial buildings by floor area (EIA CBECS 2018, Table B6), not by electric load or load factor. No detector built, no interval data analysed and no false-alarm rate, so this survey's false-alarm goal of 0.01 per facility-year is not addressed; assumes utility cooperation; off-grid generation defeats it. "A 100 kW continuous load fits within the existing population of mid-size commercial customers with refrigeration, telecom, or 24/7 operations, of which the United States has millions (U.S. Energy Information Administration 2022b). There is no externally distinguishing power signature at this scale." Posed by
| |||
| P041 | Posed | Grid sensing infrastructure and hard capsWhat grid-sensing infrastructure exists (digital relays, phasor sensors, control-system deadbands) and what hard regulatory caps bound a hidden load's size? | 8 |
Grid sensing infrastructure and hard caps that bound hidden load. Terms: PMU (phasor measurement unit): a sensor that samples grid voltage and current very precisely and time-stamped. SCADA: a utility's supervisory control-and-data system; a 'deadband' is a change too small for it to report. LEA (low-energy analog) sensor: a voltage or current sensor whose small analog output feeds a protective relay directly (IEEE C37.92). Rogowski coil: a clip-on current sensor. StatusPosed Posed by
| |||
| P042 | Started | Grid-side evasion and analog methodsWhat detection rate can grid-side methods (harmonic analysis, appliance-level load monitoring, theft detectors, time-series machine learning) keep against a load using smoothing and harmonic filters? | 3 |
Grid-side evasion and analog disaggregation methods. Terms: NILM (non-intrusive load monitoring): inferring individual loads from a single meter's total. harmonic state estimation: using the distortions a load imposes on AC power to identify it. StatusStarted Measured: a production training trace on DGX-H100 racks shows a large power swing recurring about once a second or slower. Built: GPU power smoothing in NVIDIA GB200 (ramp limits and a power floor), tested on hardware with a synthetic benchmark and simulated on the production trace (10.5% extra energy at a floor of 90% of rated power); rack batteries are modelled. These features exist for grid safety, and they also remove the rhythm a meter would look for; no detector has been tested against them. Dong et al. (2013) show that any meter-based classifier falls to chance once the hidden load's average trace matches the cover load's. A simulation study (Kimpson, Baker & Graham, October 2026) tracks the training rhythm through eight concealment strategies, but battery smoothing defeats it unless a second meter sits between the battery and the chips. The measured results so far point against meter-level detection; nothing shows every method fails. This paper proposes a multi-faceted approach to mitigate these risks, grounded in real-world production telemetry. Posed by
| |||
| P043 | Posed | Off-grid and co-located concealmentWhat is the largest compute island that evades grid-side observables via self-generation or behind-the-meter co-location, and can it be flagged by imagery or fuel records? | 4 |
Off-grid and behind-the-meter concealment of compute. StatusPosed Posed by
| |||
| P044 | Analyzed | Remote sensing of facilitiesCan remote sensing (optical, thermal, radar, seismic) estimate a data centre's power use or find undeclared sites, within what error and at what false-positive rate? | 8 |
Remote sensing (EO, thermal, SAR, seismic) of data centres. Terms: EO: electro-optical (visible-light imagery). SAR: synthetic-aperture radar (all-weather radar imaging). MWIR: mid-wave infrared (a thermal band). MASINT: measurement-and-signature intelligence (detection from physical signatures). StatusAnalyzed A desk analysis of the visual signature by facility size: cooling-tower arrays (economic above 5–10 MW) on a 2,000 m² footprint are anomalous against EIA building statistics; at 1 MW air-cooled chillers and at 100 kW a container look like ordinary commercial buildings. 'Find undeclared sites' clause only, with no false-positive rate and no search without prior candidates; the 'estimate power use within what error' clause has no held work; thermal, radar and seismic clauses are treated qualitatively (seismic sensing 'a confirmation tool, not a search tool'; vibration dampening could defeat it). The classifier result the authors cite (Clymer 2025: all facilities above 10 MW found, all four below missed) is not held and is not evidence here. "Satellite surveillance requires no cooperation from the target state, making it deployable unilaterally. It is most useful as a first filter for facilities at 10,000 H100-equivalents and above, with corroboration from other measures." Posed by
| |||
| P045 | Analyzed | Detection thresholds for undeclared facilitiesWhat is the smallest undeclared facility that combined national technical means and supply-chain tracking can detect with high probability and no prior knowledge of its location? | 6 |
Smallest undeclared facility that can be detected. StatusAnalyzed A literature-based estimate assuming the facility is not deliberately hidden; the same source says that with active concealment facilities of 10,000 to 100,000 H100-equivalents cannot be ruled out, that the 10,000 figure rests on a single open analysis and 'does not establish a lower bound on the smallest detectable facility', and that the real bound depends on what intelligence agencies achieve. No detection probability is stated, so this survey's detection goal of 90% is not shown. Supply-chain clause: a desk bound of 'on the order of a million chips unaccounted for' from Epoch's chip tracing. Nothing built or tested; the underlying classifier (Clymer 2025) is not held. "In conclusion, facilities at 10,000 H100-equivalents are still mostly detectable, primarily through satellite surveillance. Combining methods that are independent of each other raises the detection probability further. Below this scale, distinguishability drops sharply and detection becomes much harder." Posed by
| |||
| P046 | Analyzed | Detecting and attributing underwater data centresCan an underwater data centre built for evasion be detected, and told apart from commercial subsea infrastructure (commercial underwater data centres, offshore wind, cables)? | 5 |
Detecting an underwater data centre built for evasion (heat plume, acoustic and power signatures, cable installations) and telling it apart from commercial subsea infrastructure. Terms: UDC (underwater data centre): a sealed data centre placed on the seabed and cooled by the surrounding water. StatusAnalyzed A desk analysis (order-of-magnitude physics and logistics) of thermal, acoustic, optical and SAR detection of a 150 to 200 MW, 100,000-H100-equivalent underwater build, finding construction and maintenance the detectable phases; attribution against offshore wind, platforms and ships' generators is unsolved ('detection is not the same as attribution'); no field measurement, no sensor trial. Thermal detection of an operational pod is unlikely outside shallow, calm water; acoustic detection is marginally more effective, but faces limitations in attribution; and optical/SAR monitoring is most powerful during construction and maintenance Raised in: Baker, Kulp, Marks, Brundage & Heim (2025), Verifying International Agreements on AI: Six Layers of Verification for Rules on Large-Scale AI Development and Deployment (RAND WR-A4077-1); Wasil, Reed, Miller & Barnett (2024), Verification Methods for International AI Agreements; Teague, Rajmohan & Muehlhaeuser (2026), Could Underwater Data Centers Pose a Risk to AI Treaty Verification?, arXiv 2609.18824. | |||
| P047 | Analyzed | Completeness: showing that declarations cover all computeCan a state show, to a stated confidence, that it has no compute, stockpiles or facilities outside its declarations? | 6 |
Whether a state can show that its declarations cover essentially all of its compute, leaving no hidden facilities, stockpiles, parallel infrastructure or exit route, and to what confidence. StatusAnalyzed A proposed procedure only: mutual declaration of existing, in-transit and in-production compute, a supply-chain records audit with inspectors counting final chips, comparison against each side's intelligence, then extension to third nations; no confidence level, no pilot, and the plan itself says unknown compute cannot be brought under verification. Facilities clause: Koopmanschap & Barten's desk analysis finds that with active concealment facilities of 10,000 to 100,000 H100-equivalents cannot be ruled out. Stockpile and secret-production clauses: their rough bound of about a million untraced chips. Exit routes and smuggling: stated limits only (Harack et al., Aarne & Petrie). "Both the US and China should then be able to carry out an audit and inspection of each other’s declaration, and compare it against their private intelligence efforts (to the extent they properly prepared this) to resolve their uncertainties over how much compute each other have." Posed by
| |||
| P048 | Analyzed | Monitoring data sources and intelligenceCan non-technical data sources (payment and customer-identity records, financial intelligence, cloud metadata) identify large AI-chip rentals or purchases, and how reliably? | 2 |
Non-technical monitoring data sources and intelligence. Terms: KYC (know-your-customer): the customer-identity records firms must collect; used here as a non-technical signal of who is renting or buying large amounts of AI compute. StatusAnalyzed Design and feasibility argument: compute providers as KYC-style intermediaries using billing records and non-confidential metadata to classify large training and inference workloads and account for compute; no recall or precision for identifying chip rentals is measured, and detailed workload properties are said to need confidential computing. Ansari notes metadata collection stops at individual providers (no cross-provider or on-premises coverage); Leklem & Maheshwari propose a monitoring network fed by cloud API patterns and export data; Scher's 'might also use financial intelligence' is a one-line suggestion. Payment-record and financial-intelligence methods have no analysis. Several of these techniques could be used in real-time to establish whether a customer is engaging in an activity that might be subject to regulatory oversight. If compute providers were to implement these techniques, they could likely distinguish whether their customers are engaging in activities such as training a large model, or engaging in inference at scale, as well as quantify the amount of compute consumed by large workloads Posed by
| |||
Material accountancy, sampling, sequential detection and inspection games 6 problems
| ID | Status | Question | Sub-qs |
|---|---|---|---|
| P049 | Analyzed | Bounding unmonitored and undeclared computeGiven stock accounting of accelerators, what is the tightest bound on undeclared compute, and what is the smallest undeclared training run the regime can detect? | 19 |
Stock accounting of accelerators and the resulting bound on undeclared compute and on what the regime can detect. StatusAnalyzed A desk estimate from published production and ownership-tracing figures: roughly a million chips untraced, so an approximate upper bound on the largest hidden cluster; the authors call the exact lower bound hard to estimate from public information. No regime stock audit has been run. The detectable-run clause has only models: Shavit's chip-sampling analysis (unmonitored compute sets an implicit lower bound on detectable run size), Dean's AI 2040 model estimate of about 0.5% of world compute divertible, and the same paper's 10 MW (about 10,000 H100-equivalent) satellite threshold, all paper-only. This survey's run size of 10^24 FLOP is not addressed. "As noted earlier, 15.7 million of the 20.2 million chips produced can be traced to the owning company. Much of the remaining 4.5 million can likely also be traced, leaving on the order of a million chips unaccounted for. This yields an approximate bound on the size of the largest hidden cluster." Posed by
| |||
| P050 | Posed | Verifier test secrecyDo secret verifier tests catch spoofs tuned to pass the disclosed tests, and can that secrecy survive a state adversary? | 1 |
Whether secret verifier tests resist spoofs tuned to disclosed tests. StatusPosed Posed by
| |||
| P051 | Analyzed | Sampling procedures and sampling randomnessCan an adversary who controls or chooses sampling seeds encode exfiltration bits or evade detection without raising the false-positive rate? | 4 |
Sampling procedures and seed control as a verification gap. StatusAnalyzed A written analysis of the seed-control channel and the design requirement that closes it (a trusted cryptographic RNG bound to the request, not choosable or biasable by the adversary); no bit count is measured and no seed-controlling adversary is tested. Rinberg et al.'s measured exfiltration bounds (Llama and Qwen) assume trusted randomness, so they do not bear on this headline. Amodo's Stage 1 prototype records that its seed is set by the prover and should be verifier-defined, a documented gap. Speculative decoding and non-Gumbel samplers are unevaluated (DiFR). "If the adversary can manipulate randomness, they can transmit information by selecting seeds/samples while leaving per-step likelihoods within normal ranges. In this regime the verifier’s guarantees degrade sharply: many exfiltration strategies become indistinguishable from benign stochasticity, and a malicious provider could even watermark or backdoor outputs via RNG tampering while remaining white-box undetectable [GKVZ24, CG24]." Posed by
| |||
| P052 | Started | Sampled audit and recomputation ratesWhat audit and recomputation sampling rate gives a high probability of catching a violation, given that recomputation is slow? | 3 |
Sampling rates for audit and recomputation. Terms: π: the fraction of layer blocks the client samples. N: the total number of blocks. k: in Jin et al., the number of blocks an adversary alters; in Amodo's scheme, the number of steps one recomputation takes. p: the chance that a given step is recomputed. StatusStarted For this TEE-based block-commitment scheme, sampling a fraction of layer blocks gives detection probability about 1 − (1 − π)^k against an adversary touching k blocks, implemented in CUDA pipelines and evaluated on Llama-3.1-8B across attack types and block configurations; a source's own system. The slow-recomputation trade-off is only analytic: Amodo's example scheme derives a per-step recompute probability of about 1/k when recomputation takes k steps (p ≈ 0.001 for k = 1000) and notes better strategies may exist. No regime-level audit rate against this survey's detection goal of 90%. "Even modest verification efforts achieve substantial detection rates. For instance, if an adversary compromises 10% of blocks (k = 100) out of N = 1000 and the client samples only π = 1% of blocks, the detection probability reaches Pdetect ≈ 65%." Posed by
| |||
| P053 | Posed | Losses, outages and deniabilityWhat is the baseline rate of accelerators lost/unaccounted in audited fleets, and what rule separates genuine losses from strategically labelled outages? | 4 |
Chip losses, outages and plausible deniability. StatusPosed Posed by
| |||
| P054 | Analyzed | Inspection-game modelling of the regimeWhat do inspection-game models say the regime's optimal inspection rate and detectable-cheat parameters are, and how do they shift as thresholds fall over time? | 7 |
Inspection-game models of the regime and their parameters. Terms: inspection game: a mathematical model of an inspector versus a party that may cheat. In Avenhaus & Canty's notation, d is the inspected party's incentive to cheat and b the cost to it of being caught; the d/b (gain-to-sanction) ratio compares the two. StatusAnalyzed A game-theoretic model of a cloud provider detecting threshold evasion, with no empirical data: it locates a detection gap at a fixed FLOP band under FLOP-only reporting and says more models will fill it as software improves; it derives no inspection rate, no gain-to-sanction ratio and no treaty-regime equilibrium. Shavit (2023) section 3.2 gives a closed-form samples-per-year model (a sampling model, not a game). The inspection-game literature held in the collection (Avenhaus and co-authors) models nuclear safeguards, not this regime. The 'thresholds fall over time' clause has no model. "We analyzed a simple model of evasion and showed that a detection gap exists in the model and contains AI trained on biological sequence data." Raised in: Harack et al. (2025), Verification for International AI Governance (Oxford AIGI); Shavit (2023), What Does It Take to Catch a Chinchilla? Verifying Rules on Large-Scale Neural Network Training via Compute Monitoring; Canary Institute research; Daley (2026), Proving the Negative: Three Warnings from Arms Control for the Age of AI (War on the Rocks); Moon, Vedula, Geneson & Bar-on (2025), Strategies and Detection Gaps in a Game-Theoretic Model of Compute Governance (RAND RRA3686-1); Amodo (2026), Example Schemes for Verifying High-Stakes AI Agreements. | |||
Roots of trust, secure boot, firmware, TEEs and memory attestation 9 problems
| ID | Status | Question | Sub-qs |
|---|---|---|---|
| P055 | Posed | Keys and hardware roots of trustWhat lab cost and time are needed to extract an accelerator's device/attestation key, and can keys be provisioned so neither party holds a copy? | 9 |
Extraction resistance and provisioning of device keys and roots of trust. Terms: root of trust (RoT): a small, highly trusted hardware component that the rest of a system's security is built on. HSM (hardware security module): a tamper-resistant device that stores and uses cryptographic keys. Caliptra, OpenTitan: open hardware root-of-trust designs. StatusPosed Posed by
| |||
| P056 | Analyzed | Secure boot and firmware integrityCan an untrusting verifier confirm an accelerator runs only signed, non-rolled-back, auditable firmware, and has secure boot been bypassed in published work? | 8 |
Secure boot, anti-rollback and auditability of accelerator firmware. StatusAnalyzed Design options only: jointly created firmware reference measurements, and listed ways to attest that verification firmware stays present (secure boot made impossible to disable, a secure counter, randomly timed low-latency attestation); Shavit 2023 proposes signed-firmware enforcement, anti-rollback and early remote attestation of updates. Nothing built or tested. Bypass clause at a lower stage: no held document reports a secure-boot bypass of a data-centre GPU; flexHEG Part 2 describes voltage-glitch attacks on secure boot in general from a source not held; Skorobogatov & Woods 2012 broke an FPGA's protections, not an accelerator's. H100 secure boot and anti-rollback features are reported second-hand (Petrie 2024; Scher & Thiergart) from NVIDIA documents not held. To obtain reliable reference measurements when the hardware designer’s measurements are untrusted, without gaining access to private information such as the Prover’s code, one option may be for the Verifier and Prover to jointly create reference measurements by emulating a compliant bootup on a confidentiality-preserving technology (e.g., Confidential Computing). Posed by
| |||
| P057 | Analyzed | Governance of mechanism firmware and rule updatesCan updates to governance firmware/rulesets be signed, rollback-protected, multi-party-controlled, and constrained so no single party can change the rules? | 4 |
How updates to governance firmware/rulesets are signed, forced, rolled back or constrained, and who may change the rules. StatusAnalyzed Design discussion only (flexHEG Part III Appendix A; Part II 'Guarantee Update Process'): forced updates through limited lifetimes or licences that carry a minimum version number; approver-based or k-of-n signing of updates; approver signing keys held in a guarantee processor that serial-numbers what it signs, to prevent secret double-signing of lifetime extensions ('a topic for further research'). Nothing built or independently reviewed. Rollback protection is assumed, not shown. The clause 'constrained so no single party can change the rules' is at a lower stage: the same authors say rules about future rules are 'difficult to articulate', a baseline ruleset would be 'technically very challenging', and there is 'no clear mechanism' to ensure neutral rules. Required updates would be implemented by having each update have a limited lifetime, tracked by an onboard clock. … If no new update has been installed before the lifetime runs out, the device locks down. Posed by
| |||
| P058 | Started | GPU TEE security evidenceWhat independent, non-vendor security evidence exists for GPU secure enclaves (TEEs) against physical and side-channel attack, and how many in-scope breaks have been found? | 3 |
Independent security evidence for GPU TEEs, incl. physical and side-channel attacks. Terms: TEE (trusted execution environment): an isolated, hardware-protected area of a processor whose code and data the machine's owner cannot read or alter. StatusStarted One independent test of NVIDIA GPU confidential computing under physical attack, and it found a break. Academic researchers built a DDR5 memory-bus probe for under $1,000, extracted Intel's attestation key from a fully trusted Xeon in about 15 minutes, and ran a workload on an RTX 3060 outside any secure enclave while passing H100 confidential-computing attestation relayed from a rented H100, because the GPU attestation is not tied to a specific VM. The break goes through the CPU; the H100's own memory and enclave were not attacked, and both vendors call physical attacks out of scope. An independent reverse-engineering of H100 confidential computing (Gu et al. 2025) finds the boot chain and attestation sound within NVIDIA's threat model, but host-visible plaintext command metadata that allows partial reordering or dropping of firmware calls, unprotected command queues, a transfer-size timing channel and still-readable registers: leaks and partial integrity loss, not a break. No physical attack on a data-centre GPU's HBM is held. A Rowhammer attack on workstation GPUs (Lin et al. 2026, GPUBreach) is a cross-reference: its authors list confidential-computing mode as untested. proves the CPU interfaces directly with a genuine CC GPU. However, the attestation does not identify what the GPU is running. Posed by
| |||
| P059 | Started | GPU confidential computing and link encryption: availability and overheadWhich shipping accelerators support multi-GPU confidential computing and encrypted links, at what throughput cost and with what key handling? | 13 |
Which accelerators ship multi-GPU confidential computing / encrypted NVLink and links, at what throughput cost, with what binding and key handling. Terms: confidential computing: keeping data encrypted even while a chip is processing it. NVLink: Nvidia's high-speed link between GPUs. AES-GCM: a standard method for encrypting and authenticating data in transit. StatusStarted Measured on one Huawei Ascend 910A accelerator, a vendor research prototype of firmware-only confidential computing: 0.028-0.91% added inference latency. Single device and inference only; no multi-device or training overhead has been published. A workshop participant reports an unpublished characterisation of training overhead from 2 to 1,024 GPUs; it is not counted, since no numbers are public. Availability: in January 2026 only the H100 offered working confidential computing; cluster-scale confidential computing on Blackwell at near-unencrypted speed is a vendor claim. Link encryption: roughly 3% of compute to encrypt NVLink traffic is a desk estimate. introduces minimal performance loss during the inference pass(0.91% in GPT2-neo125M and 0.028% in Llama3-Chat-QA-1.5-8B model Posed by
| |||
| P060 | Analyzed | Weight snapshots on acceleratorsCan an accelerator take attested, tamper-evident weight/memory snapshots bound to chip state at trigger time, at what overhead and with what storage cost? | 10 |
Attested, tamper-evident snapshots of weights/memory on chip: support, overhead, timing binding, locating weights, storage. Terms: HBM (high-bandwidth memory): the fast memory stacked next to an accelerator where model weights and activations are held. StatusAnalyzed Design only: random-time, firmware-raised HBM snapshots hashed on chip and kept in firmware-only non-volatile memory or signed to a verifier-trusted server; overhead and storage cost discussed qualitatively (synchronised chips hang during a snapshot; terabyte-scale checkpoints). Assumes hardware-security-module features that Harack et al. call 'speculative' and 'not yet available'. Harack Appendix L.4 adds a design sketch for a timed weight-residency challenge of about 50 ms, 'an area for future work'. Nothing built. Related built work: Monfared et al. 2026 measured a gap above 350 ms between HBM-resident and host-RAM-resident 60 GB challenges on an H100, a residency probe rather than an attested snapshot, so a cross-reference. A chip’s weight-shards can be captured by snapshotting the chip’s high-bandwidth memory, which contains the weights. These snapshots must be saved at random times, as otherwise the Prover can swap out the true weights for a fake set of weights right before the scheduled logging. Posed by
| |||
| P061 | Started | Verifiable memory wipes and memory challengesCan a node's HBM/DRAM/SSD and network-equipment memory be proven wiped or filled within a stated time and challenge latency, and how much non-wipeable storage remains? | 10 |
Proving HBM/DRAM/SSD and network-equipment memory is wiped or filled: wipe time, non-wipeable storage, challenge latency and soundness. StatusStarted Proof-of-secure-erasure experiments on H200 hardware, the authors' own report: RAM and HBM wipe estimated at tens of minutes for a GB200 tray; disk wipe about 24 hours for 15 TB, extrapolated from a measured 1 TB NVMe write; challenge-phase time not yet computed; non-wipeable memory targeted below 1 GiB but not yet quantified; network-switch memory listed as open. Challenge clause: Monfared et al. 2026 measured a gap above 350 ms between HBM-resident and host-RAM-resident 60 GB challenges on an H100, with no false-alarm rates; Reuter 2026's roughly 2x margin against host-DRAM caching is an analysis; Bursuc et al. 2024 give provably secure proofs of erasure with round-trip-time bounds (theory). Cankaya's v0.2 draft (dated) says a network-level timing probe telling one node's DRAM from another's 'has not yet been demonstrated'. Both clauses sit in one memory-verification community; not every clause answered. Writing a full 1 TB gave a final average write speed of 160 MiB/s — much lower than the advertised rate — taking ~1h 35 minutes. Posed by
| |||
| P062 | Analyzed | Integrity and attack surface of the verification apparatusCan the verifier-side infrastructure itself (recomputation servers, taps, sensors) be trusted and resist a state attacker, and how much attack surface does it add? | 4 |
Whether verifier-side infrastructure (recomputation servers, taps, sensors, the regime as a whole) can itself be trusted and resist state attackers. StatusAnalyzed Design only, from a self-deprecated working draft (June 2026, since rewritten; treat as dated): a verifier-side auditing environment built from three independently sourced trusted computing bases whose outputs are compared before leaving an air-gap, plus an instructor-executor split with formally verified logic for the simple parts. Nothing built; no estimate of added attack surface. Recomputation-server hardening has no work (AI 2040: 'needs significant research'; Amodo: 'has not been explored'). Comer et al. 2026 (RAND) design a secure inference facility against nation-state attackers, but prover-side, so a cross-reference. At least three distinct TCBs (different supply chains, hardware and software. One trusted by the prover, the verifier and a third-party) make up three small clusters, all inside the prover’s datacenter and inside their own air-gap. Posed by
| |||
| P063 | Analyzed | Composing trust across parties and componentsCan a mutually secure verification system be built from unilaterally trusted components, and can the combination of trust roots be proved sound? | 2 |
Composing trust from unilaterally trusted components. StatusAnalyzed Design and a stated soundness condition only, from a self-deprecated working draft (June 2026; treat as dated): a mutually secure auditing environment from three unilaterally trusted, independently sourced trusted computing bases; the February 2026 workshop report says the approach is 'designed to work with great power adversaries without trusting either side's chips', with specifics 'TBA'. Nothing built. The 'proved sound' clause has no work: AI 2040 calls proving the combination sound 'the open problem'. Hausenloy & Li's 'never trustless, only low-trust' is a position. The security of redundant cross-checking hinges on the different devices not having the same vulnerabilities, which means they need to be built from independently sourced components, and not share (too much) code or architectural similarity in general. Posed by
| |||
On-chip enforcement, licences, metering, trusted time and on-device logging 13 problems
| ID | Status | Question | Sub-qs |
|---|---|---|---|
| P064 | Analyzed | Robustness of on-chip governance mechanismsAre on-chip governance mechanisms robust against physical attack and hard to circumvent? | 1 |
Whether on-chip governance mechanisms are robust against physical attack, confidential enough, and hard to circumvent. StatusAnalyzed Designs only. Petrie (2025) designs an off-switch built from thousands of redundant security blocks per chip and evaluates attack vectors on paper; Petrie's firmware offline licensing (2024) is an earlier design. No on-chip governance mechanism has been fabricated and tested against physical attack. The flexHEG authors mention a power-measurement prototype in one sentence, with no results published; a tamper-responsive enclosure prototype tested against probes is cited in flexHEG Part 2, but its own report has not been located. This massively redundant architecture is designed to prevent unauthorized chip use, even against sophisticated physical attacks. Raised in: Reuel, Bucknall et al. (2025), Open Problems in Technical AI Governance, compute sections; Aarne & Petrie (2025), International Security Applications of Flexible Hardware-Enabled Guarantees (flexHEG Part 3); Aarne, Fist & Withers (2024), Secure, Governable Chips (CNAS); Ansari (2026), Hardware-Level Governance of AI Compute: A Feasibility Taxonomy for Regulatory Compliance and Treaty Verification; Kulp et al. (2024), Hardware-Enabled Governance Mechanisms (RAND WR-A3056-1); Wasil, Reed, Miller & Barnett (2024), Verification Methods for International AI Agreements; Proofworks, Living Verification Documents. | |||
| P065 | Analyzed | Time to deploy on-chip governance mechanismsHow long would on-chip mechanisms take to design, test and reach the installed base? | 1 |
How long on-chip mechanisms take to design, test externally and reach a stated share of the installed base. StatusAnalyzed A modelled estimate: 3.7-7.9 years for chips with built-in flexHEG hardware to displace older accelerators in frontier AI development, counted from when the manufacturer starts work. Other published estimates: 'at least a few years' (Harack et al. 2025); months for firmware-only mechanisms and 18 months to 4 years for silicon changes (Aarne, Fist & Withers 2024). No estimate has been checked against an actual deployment. this estimate predicts 3.7 - 7.9 years for flexHEG with integrated hardware to displace non-flexHEG accelerators in frontier AI development Raised in: Harack et al. (2025), Verification for International AI Governance (Oxford AIGI); Scher & Thiergart (2024), Mechanisms to Verify International Agreements About AI Development (MIRI); Petrie & Aarne (2025), Technical Options for Flexible Hardware-Enabled Guarantees (flexHEG Part 2); Aarne & Petrie (2025), International Security Applications of Flexible Hardware-Enabled Guarantees (flexHEG Part 3); Aarne, Fist & Withers (2024), Secure, Governable Chips (CNAS); Cankaya (2026), Compute Verification FAQs; Ansari (2026), Hardware-Level Governance of AI Compute: A Feasibility Taxonomy for Regulatory Compliance and Treaty Verification; Kulp et al. (2024), Hardware-Enabled Governance Mechanisms (RAND WR-A3056-1); Wasil, Reed, Miller & Barnett (2024), Verification Methods for International AI Agreements; Peigné-Lefebvre, Nguyen & Wang (2026), Zero Knowledge Verification for Frontier AI Training Is Possible, arXiv 2606.05433. | |||
| P066 | Analyzed | Guarantee processor and verification co-hardwareCan a guarantee processor (verification co-hardware) be interposed on accelerator I/O at acceptable overhead, with binding that detects its removal? | 10 |
Guarantee processors and verification co-hardware: overhead, binding, feature set. Terms: guarantee processor (verification co-hardware): a separate, trusted chip placed alongside an accelerator to watch its inputs and outputs and enforce or attest rules. StatusAnalyzed A design, not a build: the flexHEG Guarantee Processor plus Interlock sitting on the accelerator data path, with four candidate placements and a qualitative latency/throughput trade-off (requirements rise the closer the interlock is to the die). Overhead clause: not analysed anywhere; Ansari 2026 says the performance cost of mediating all chip I/O is 'not yet addressed in the published literature'. Binding clause: only qualitative statements (retrofitted hardware 'may be easier for an attacker to decouple'; flexHEG Part 3 poses how the 'brain' could be integrated so it cannot be easily swapped out as an open question); no pairing scheme, no detection rate. The flexHEG team's power-measurement prototype is a retrofit guarantee processor that makes external measurements and does not sit on the I/O path, so it is a cross-reference, not evidence for this headline. The Interlock could be implemented in several ways, including 1) as an IP block within the accelerator die, 2) as a chiplet sitting near the accelerator die, 3) as a part of high bandwidth memory (HBM), or 4) as a separate component that operates as a network relay or switch. Posed by
| |||
| P067 | Analyzed | Chip operating licences and off-switches: the on-chip mechanismCan a signed, time-limited operating licence or remote off-switch be built into accelerators so that it resists downgrade, bypass and stockpiling? | 9 |
Whether a signed, time-limited operating licence or remote off-switch can be built on shipping accelerators, its renewal and anti-stockpiling design, and its resistance to firmware downgrade, bypass and physical key extraction. StatusAnalyzed Designs only, nothing built: Petrie's on-die design of about 10,000 redundant licence-checking security blocks with an attack-and-countermeasure table (logic flaws, licence reuse, glitching, focused-ion-beam edits, key extraction); RAND's offline licensing with metered compute budgets (Kulp et al. 2024); CNAS's timed licences as the only robust remote disable (Aarne, Fist & Withers 2024). No shipping accelerator enforces a signed time-limited licence (Harack et al. 2025: 'not yet available'). Downgrade and bypass resistance: asserted, not tested; no red-team of licensing firmware on an AI GPU. Stockpiling: discussed as a risk (cycle-counter banking, powered-off reserve chips) with design responses, nothing shown. Related but not counted as a tested failure of this family: CNAS reports that Nvidia's driver-level hash-rate limiter on consumer GPUs was fully circumvented about a year after release, a different mechanism family reported second-hand. In this paper, we propose a significantly more secure design that would embed thousands of hardware security blocks into upcoming generations of AI chips. Each security block would individually check for up-to-date usage authorization, and block essential chip operations if this approval is not received. Posed by
| |||
| P068 | Analyzed | Licence issuance and key governanceWho should issue operating licences and hold the signing keys, and how can issuance and emergency-disable powers be made abuse-resistant? | 5 |
Who issues operating licences and holds the signing keys (criteria, k-of-n signer sets across jurisdictions, bootstrapping, liveness), and how issuance and emergency-disable powers are kept abuse-resistant. Terms: k of n (k-of-n): a threshold scheme in which an action needs the approval of any k of the n parties, so it keeps working after up to n−k keys are lost. StatusAnalyzed A protocol sketch with pseudocode: k-of-n multilateral signing of licence and guarantee updates, graded quorums by update type, and signer-set changes by update. No prototype (Ansari 2026: 'No prototype exists'), and the open parts are stated by the proposers themselves: who the licence authority is, how the signer set is bootstrapped, what k and n should be, timeout handling for unresponsive signatories, proofs of non-issuance ('would probably require hardware-backed evidence from all devices that contain the private key'), and the blast radius of a stolen signing key. Abuse resistance of emergency disable: Baker et al. 2025 judge it not to need significant R&D because proposals exist; that is a readiness opinion, not evidence. The update process could require that the update (or lack of update) be multilaterally approved by k of n parties, so that no single actor could unilaterally impose excessive restrictions (as discussed for an international context in Part III). The pseudocode in algorithm 1 shows how multi-party update approval could be implemented. Posed by
| |||
| P069 | Analyzed | Hardware enforcement of permitted useCan chips be restricted to approved code, inference-only, or bounded training, and what throughput penalty does the best circumvention attack achieve? | 10 |
Restricting chips to approved code, inference-only operation or bounded training, the cost of circumventing such restrictions, and the time to convert deployed compute to verified inference-only operation. StatusAnalyzed Designs only: RAND's licence-metered throttle (bounded training by operation budget, with a hardware cutoff sketch) and CNAS's on-chip restriction to approved code or models; Scher & Thiergart list inference-only chips and an operation disallow-list as building blocks and call them possibly 'extremely difficult or effectively impossible'. No prototype has refused a non-compliant job in a test (Harack et al. 2025). The circumvention clause has no work: no measured throughput penalty for the best attack, and RAND itself notes a clock-slowing throttle 'might be easy to circumvent with a cheap FIB edit'. Stated limit, not a failure: software-only mechanisms can verify use but not enforce it (Petrie & Aarne 2025). When the counter for a functional unit (such as a floating-point multiplier block) exceeds the threshold set by the license, a multiplexer placed before the output of the functional unit can select an alternative output value of zero, ignoring the actual computation results. Posed by
| |||
| P070 | Analyzed | Compute metering and trusted timeCan an on-chip compute and activity meter and a trusted timer count accurately and resist clock, fault and counter-manipulation attacks? | 10 |
On-chip compute meters, counters and trusted time. StatusAnalyzed Design only: a meter built on existing performance counters with licence-set caps, plus proposed protections (monotonic counters, lockstep redundancy, shielding against fault injection, sanity checks between counters). Nothing fabricated or measured; Ansari 2026 records 'the absence of functional prototypes', and current counters are 'accessible via firmware interfaces and were designed for debugging rather than tamper-resistant governance'. Trusted-time clause: design-level only; CNAS proposes counting clock cycles instead of time because tamper-proof time 'generally is considered very difficult', and RAND says there is 'no airtight defense' against clock attacks. Fault-injection resistance and sanity-check false-alarm rates: no work. Since these values are already collected in hardware, additional logic could be added to set a maximum value for each counter, determined by a license. Posed by
| |||
| P071 | Posed | Detecting fabricated logsWhich log-fabrication attacks are hardest to detect, and how much cross-device correlation is needed to catch them? | 4 |
Detecting fabricated or incomplete activity logs from an operator with root and physical access: the hardest attacks, cross-device correlation, tamper evidence, logging standards and binding data to a chip. StatusPosed Idea stage only: a proposed red-team/blue-team exercise to test whether cross-device log analysis catches fabricated logs, not yet run; no design for detecting fabrication. Posed by
| |||
| P072 | Started | Challenge-response timing checksCan challenge-response timing checks detect precomputed or decoy work as a cheaper alternative to recomputation? | 1 |
Challenge-response timing checks as a cheaper alternative to recomputation for detecting precomputed or decoy work, against an operator with root and physical access. StatusStarted Four verifier-issued timing probes implemented and run on NVIDIA T4 and H100 GPUs on cloud VMs: proof-of-work, parallel verifiable delay functions, tensor-core GEMM puzzles and a randomly timed VRAM-residency check whose keyed Argon2id pass resists parallel precomputation; solve times separate idle from contended GPUs and HBM-resident from host-resident data (60 GB challenge, more than 350 ms apart). The authors call the probes heuristic indicators, give no false-alarm or miss rates, test no adaptive evader, allow answering on idle identical GPUs, and do not compare cost with recomputation; precomputed or decoy ML work as such is not the tested attack. Cankaya (2025) reports that in his own experiments timing revealed resource contention from hidden parallel work (reported, not published in detail). the timing gap between the two modes exceeds 350 ms, making them trivial to distinguish. Raised in: Cankaya (2025), Catching Misreporting About ML Hardware Use by Turning Noise into Signal; Cankaya (2026), A System Overview for Near-Term, Low-Trust AI Compute Verification (v0.2); Dean, Verification Plan (AI 2040). | |||
| P073 | Analyzed | On-device compute-graph logging and guarantee checkingCan on-device compute-graph logging run on a multi-node job with verified completeness and low overhead, and how many falsified operations go undetected? | 7 |
On-device compute-graph logging and its falsification. StatusAnalyzed Design only: compute-graph declaration to a guarantee processor, static or dynamic, with a proposed description language and kernel translator, and a back-of-envelope log rate (about 10 KB/s per accelerator, which the authors say 'does not scale well with cluster size' when stored locally). Nothing run on a multi-node job; completeness and overhead unmeasured; falsified-operation detection posed as research questions, not tested. Related built work that is not this mechanism: ShadowScope (Almusaddar et al. 2025) validates single-GPU kernel execution against golden performance-counter traces with a non-adaptive attacker; the headline names multi-node logging, so that single-GPU result is a cross-reference. Alternatively, a description language specifically for flexHEG compute graphs could be developed along with tools to translate accelerator kernels into this language. Posed by
| |||
| P074 | Analyzed | Guarantee-checking logic and its limitsWhat are the limits of automated guarantee-checking of workloads, and what recomputation fraction is needed to catch a falsified kernel? | 5 |
Limits of automated guarantee checking of workloads. StatusAnalyzed Proposals and analysis only: random partial recomputation by the guarantee processor, local LLM-based checks, and the authors' own argument that fully automatic on-device compliance checking 'appears unlikely to be feasible'; Reuter, Marks & Baker (2026) give a formal framing (a policy over an algebraic graph; 'functional exclusion' left open). No recomputation fraction has been derived or measured for any detection probability; no automated checker has been run on camouflaged logs. Stated limits: static graphs cannot express mixture-of-experts routing; the relevance of Rice's theorem is 'unclear'. a general-purpose way to verify each atomic computation is to use the guarantee processor to randomly re-compute a portion of each result Posed by
| |||
| P075 | Started | Network interlock performanceCan an inline network interlock enforce a signed list of allowed peers at full line rate (800 Gb/s to 1.6 Tb/s)? | 3 |
Whether an inline network interlock can enforce a signed peer allow-list at full line rate: latency, firmware updates, key setup, in-network-compute switches, NVLink traffic and DMA spoofing. StatusStarted Built, not measured: Amodo report a working prototype in which a central controller leases traffic quotas to NVIDIA BlueField network cards that enforce them, and hardware-accelerated encryption between nodes, with keys the controller can revoke. Membership by key is a form of the allowed-pairs list the problem asks about, but no peer allow-list has been built as such, and no line-rate or latency figures are published. The full interlock design is in flexHEG Part 2 (Petrie & Aarne 2025), with the network card as the retrofit location; encrypting all NVLink traffic is estimated at about 3% of H100 compute (a desk estimate). Got hardware-accelerated encryption working Posed by
| |||
| P076 | Posed | Trusting a network card from an adversarial vendorCan a commodity NIC serve as the trust anchor if its vendor is adversarial? | 1 |
Whether a commodity network card can serve as the verification trust anchor when its vendor is adversarial. StatusPosed Posed by
| |||
Replay, determinism, floating-point reproducibility and compute accounting 11 problems
| ID | Status | Question | Sub-qs |
|---|---|---|---|
| P077 | Started | Cross-hardware bit-exact replay and emulationCan a computation be replayed or emulated bit-exactly on hardware other than the prover's (other SKUs, vendors, CPU), and at what slowdown? | 12 |
Replaying or emulating computations bit-exactly on hardware other than the prover's (other SKUs, vendors, CPU emulation, verification-chip variants). Terms: GEMM (general matrix multiply): the dense matrix-multiplication that dominates AI compute. MoE (mixture of experts): a model design that routes each input through only some of its sub-networks. StatusStarted Software-only CPU emulation of NVIDIA GPU inference arithmetic (L40, L40S, A100, H100 profiles), dense transformer blocks of Qwen3-4B plus one full forward pass; bit-exact against CUTLASS, and against cuBLAS through a per-SKU dispatch catalogue except nvjet kernels on H100 at short sequences. Not mixture-of-experts, not AMD, TPU or Ascend, not training. Slowdown clause: not measured; the appendix estimates CPU need and says the figures 'likely need to be multiplied by at least 10x' for reduction-tree emulation, so that clause is paper-only. Cross-hardware next-token agreement without emulation is about 85% one GPU generation apart (Hausenloy & Li 2026). The emulator matches CUTLASS bit-exactly at every tested hardware/sequence-length combination. Posed by
| |||
| P078 | Analyzed | Compute accounting: MFU, HFU and declared operationsCan a model predict achievable hardware utilisation closely enough to detect padded or undeclared-efficiency compute, at a tolerable false-alarm rate? | 5 |
MFU/HFU modelling and detecting padded or undeclared-efficiency compute. Terms: MFU / HFU (model-FLOP utilisation / hardware-FLOP utilisation): the fraction of a chip's peak compute a training run actually uses - two accounting conventions that differ in how they count recomputed work. StatusAnalyzed A proposed method with a desk analysis of its error sources: estimate achievable utilisation for a declared workload and cluster, compare with declared operations, and back it with physical signatures when an operator's efficiency exceeds the estimate. No model has been validated against measured runs, no padding or no-op detector has been built, and the false-alarm rate for honest operators is only flagged as a risk ('honest Provers [may be] unable to account for their estimated compute use'). Attacks named without defences shown: secret backend optimisation, speculative decoding (1.5-3x), Pareto-inefficient menu choices, restructuring to minimise logged operations. Cankaya's system overview proposes a resource-occupation simulation cross-compared with throughput, also paper-only. For a given workload type and cluster, optimal MFU could be coarsely approximated as the MFU known to have been achieved on similar workloads and clusters, on which there is private and public data (Epoch AI, 2024b). Optimal MFU could be more precisely estimated with theoretical modeling (Erdil and Schneider-Joseph, 2024; Erdil, 2025). Still, significant error bars may remain. Posed by
| |||
| P079 | Analyzed | Independent verification of declared training-compute totalsCan any independent method - not relying on the chip's own secure enclave - verify a run's declared total compute on a real distributed run, to a stated error? | 3 |
Independent verification of a run's declared FLOP total. StatusAnalyzed Proposals only, none run on a real distributed run: power-draw times energy-efficiency accounting of total operations (RAND 2025), workload-modelling of achievable utilisation (same source), spot-check zero-knowledge proofs with running cumulative-FLOP attestations (Peigné-Lefebvre, Nguyen & Wang 2026, 'a design, not a build'), and chip-logged weight snapshots with verifier retraining (Shavit 2023). No stated error for any of them on a real run; Ansari 2026 and Peigné-Lefebvre et al. both say no independent primitive for verifying reported training compute exists. Built but not on this quantity: Abdelghafar & Kulp 2026 classify training versus inference from real single-machine telemetry and say their duration tier 'is not a proxy for statutory training-compute thresholds'; Harack's Lawfare example is a code-inspection demonstration, not compute totals. Confidential-computing certificate rows depend on the chip's own enclave, which the headline excludes, so they are related, not evidence. To verify the number of total operations done in given AI data centers in some period (from ta to tb ), one approach could be to measure AI chips’ power draw (P) over time (Watts per second), and then multiply it by the chips’ energy efficiency (OP per Watt) Posed by
| |||
| P080 | Started | Bit-exact training replayWhat fraction of training steps replay bitwise on a verifier's hardware, and what throughput penalty does deterministic training impose? | 9 |
Bit-exact replay of training steps. Terms: NCCL: Nvidia's library for exchanging data between GPUs during training. ZeRO-3: a technique that splits a model's parameters across many GPUs to save memory. StatusStarted Built and measured: single-GPU training of small models (ResNet-50, 23M parameters, from scratch; GPT-2, 117M, fine-tune) replayed bit-identically across three NVIDIA consumer and workstation GPUs (A40, Titan XP, RTX 2080 Ti), FP64 with logged rounding decisions, at about 1.2-1.4x trainer and 1.3-1.7x auditor time and 18-20 MB of logs per GPT-2 step. Not multi-GPU or distributed (all-reduce ordering unhandled), not low precision, not frontier scale. Throughput-penalty clause: one measured figure elsewhere, Peigné-Lefebvre et al. 2026 report full determinism on a 7B model on 8 H100s costing 1.6-8.2% per step and cite 23-82% overhead for deterministic FlashAttention backward. Fraction-of-steps clause at frontier scale: no work. A method for two parties, training the same model on different GPU types, to achieve identical training results by logging and sharing rounding decisions Posed by
| |||
| P081 | Started | Bit-exact inference replay (same hardware class)Do inference requests replay bit-exactly on a second machine of the same SKU at production batch sizes, and at what overhead? | 6 |
Bit-exact replay of inference on same-class hardware. StatusStarted Closest to Done on the list: two independent groups show same-SKU inference replays bit-exactly once batch composition is controlled; what is missing is one run on a second machine at production batch size. Same SKU, same software stack, no floating-point atomics in the kernels, batch size logged (one integer per forward pass); shown for Qwen2.5-7B prefill across physical cards at batch sizes 1, 2 and 4. Corroborated by an independent group with its own method: He & Thinking Machines Lab (2025) report 1,000 identical temperature-0 Qwen3-235B completions with batch-invariant kernels at about 20% matmul throughput loss, on the same server. Not shown: production batch sizes on a second machine; mixture-of-experts under tensor parallelism across machines; INT-quantised models, where Cankaya saw non-determinism in vLLM. Earlier 'bitwise reproducibility is rarely achievable in practice' (Karvonen et al. 2025) describes default stacks and predates both. The within-SKU zeros (A100-SXM4 ↔ A100-PCIe) rule out physical-card variance: yield-binned chips of the same SKU produce bit-identical outputs. Posed by
| |||
| P082 | Started | Statistical inference verification (TOPLOC / DiFR family)Do statistical inference-verification methods reliably catch cheating with few false failures on large models, and do adaptive attacks defeat them? | 7 |
Statistical inference verification (TOPLOC / Token-DiFR family). Terms: TOPLOC, Token-DiFR: named published methods that statistically check whether a model really produced the outputs it claims, without re-running the whole computation. k: in top-k sampling, the number of most likely tokens the sampler chooses among; the "simulated k = 2 bug" makes 1% of tokens a uniform pick from the top two. StatusStarted Token-DiFR re-implemented by a third party and run on a tray of H100s against non-adaptive attacks (substituted models including gpt-oss-120b versus 20b, LoRA append and append-on-trigger on Qwen2.5-7B); the best honest-versus-cheating pass-rate gap was about 25%, the threshold must be set per model, and most misses were outputs under 5 tokens. Adaptive-attacks clause: the DiFR paper (Karvonen et al. 2025) shows Token-DiFR robust to one simple adaptive attack (temperature tuned to match mean cross-entropy, which defeats cross-entropy checks); stronger adaptive attacks untested, and Amodo's own note says its testing 'only covers a fraction of possible attacks'. Models tested are well below frontier size; the smallest deviations on Qwen3-30B-A3B 'remain difficult to separate' in the DiFR paper. Our results add further evidence that Token DiFR can identify different incorrect-model attacks reliably across a small range of prompts. Posed by
| |||
| P083 | Started | Inference verification overheadHow much latency does inference verification add, and how much serving throughput does it cost? | 3 |
The latency, serving throughput, verifier compute and record size that deploying inference verification imposes. StatusStarted The authors' own measurements on Llama-2-13B (Table 6): committing overhead 0.26 ms per token, validation time 81 ms, 10 KB of provider GPU memory per token; a measured per-token latency cost, not serving throughput at production batch sizes or end-to-end latency on a deployed service. Amodo's Stage-1 prototype measures verifier GPU utilisation (about 2% versus about 53% for the prover on one H100 tray, small open-weight models) and lists latency and throughput as future work; Karvonen et al. (DiFR) measure record size (1.5 to 30 bytes per token), which is in the statement, not the headline. Time and memory overhead of generating and validating Llama-2-13B inference using established verifiable inference compared to our method. Raised in: Cankaya (2025), Catching Misreporting About ML Hardware Use by Turning Noise into Signal; Cankaya (2026), Bit-Exact AI Inference Verification Without Performance Tradeoffs, arXiv 2606.00279v2; Cankaya (2026), Compute Verification FAQs; Karvonen et al. (2025), DiFR: Inference Verification Despite Nondeterminism; Amodo (2026), An Inference Verification Prototype — Stage 1; Proofworks, Living Verification Documents. | |||
| P084 | Started | Inference verification on evolving architecturesDoes recomputation-based verification keep working as model architectures and inference techniques change? | 1 |
Whether recomputation-based inference verification keeps separating honest from substituted computation as model architectures and inference techniques change. StatusStarted Measured on real GPUs for three open-weight models (Llama-3.1 8B, Qwen3-8B and the mixture-of-experts Qwen3-30B-A3B), with A100-served and H200-verified cross-hardware noise, as of late 2025; the authors report robust detection across these architectures. Speculative decoding is not evaluated (vLLM 0.10.0 lacked it) and only a verification algorithm for one variant is proposed in an appendix; constrained decoding and future architectures are untested. The AI 2040 and Amodo calls for developers to keep recomputation algorithms in lock-step with new architectures are marked not started. These models cover a wide range of differences between models such as hidden dimensions, training distributions and broader architectural differences such as being dense or mixture-of-experts. Posed by
| |||
| P085 | Analyzed | Inference verification for closed-weight models in productionCan inference be verified for closed-weight models on a lab's production stack? | 2 |
Verifying inference for closed-weight models on a lab's own production serving stack, e.g. inside an attested enclave. StatusAnalyzed Design only: the AI 2040 and Amodo plan verifies production inference by tapping the prover's network and recomputing on a verifier-trusted server placed inside the prover's data centre, so closed weights never leave; the SITREP marks recomputation-server security 'Not on track'. Amodo asks labs to re-implement its open-weight demo on their own stacks with frontier models (not done); Karvonen et al. state DiFR cannot verify closed-weight providers (a design limit, not a test); Cankaya says proprietary-model verification would need third-party- or hardware-attested deployment (one sentence). Whilst simpler than developing a whole SL5 cluster, the recomputation server needs to sit within the prover's data centre, possibly under their physical control. Raised in: Cankaya (2026), All Hands on Deck to Build the Datacenter Lie Detector; Karvonen et al. (2025), DiFR: Inference Verification Despite Nondeterminism; Amodo (2026), AI 2040 Plan A — How You Can Help. | |||
| P086 | Analyzed | Mutually trusted verification clusterCan a jointly trusted neutral cluster be built, costed and red-teamed that replays training/inference with a measured leak bound, trusted by both parties? | 4 |
Building, trusting and operating a jointly trusted (neutral) cluster that replays or verifies training and inference. StatusAnalyzed A design sketch of a neutral jointly-trusted replay cluster (no external communication, no persistent storage, one cluster per chip type), with further designs in Harack et al. (2025) section 2.5.3 and the Proofworks mutually-trusted-compute document. Nothing built: no costed design, no two-party red team, no measured leak bound. Harack et al.: the barriers are security issues with no publicly-known robust solutions. The workshop note that one ~100-chip neutral datacenter project exists is unattributed and does not count. "Instead, the Prover and Verifier can agree on a neutral jointly-trusted cluster, which they will use to execute the verification protocol. The Verifier needs to be able to trust in the integrity of the cluster’s computation. The Prover, conversely, needs to trust that the cluster will not reveal its private data, by verifying that it cannot communicate with the outside world and has no persistent storage." Posed by
| |||
| P087 | Posed | Hardware heterogeneity and technology changeDoes any verification protocol run unmodified across accelerators from more than one vendor, and how much work does porting to a new hardware generation take? | 4 |
Hardware heterogeneity and portability of protocols. Terms: hardware heterogeneity: a regime having to work across accelerators from different vendors and generations. ICI (inter-chip interconnect): the high-speed links joining accelerators (e.g. on Google TPUs). flexHEG: a specific proposed open hardware-governance design. StatusPosed Posed by
| |||
Verifiable computation, ZK proofs, network commitments and minimal disclosure 5 problems
| ID | Status | Question | Sub-qs |
|---|---|---|---|
| P088 | Started | Proof-of-learning / proof-of-training-transcriptIs there a proof-of-learning/transcript scheme with a formal soundness bound against an adaptive spoofer, and at what verifier cost as a fraction of training? | 11 |
Proof-of-learning / proof-of-training-transcript soundness and cost. Terms: proof-of-learning (PoL) / proof-of-training-transcript (PoTT): schemes that produce evidence a specific training run actually happened as described, checkable by a verifier. StatusStarted Heuristic Proof-of-Training-Data defences (certified-randomness seed, local memorization tests) tested on GPT-2-124M and Pythia checkpoints against all attacks then published in the Proof-of-Learning literature, at about 1.3% of training compute for the verifier; language models only, single attacks only, composite and adaptive attacks and online RL untested, and the Verifier sees raw data and weights. No formal soundness bound exists: the earlier Proof-of-Learning family (Jia et al. 2021) was broken by Fang et al. (2023), a tested failure of that mechanism family; Shavit (2023) gives the formal PoTT definition and leaves provably secure methods to future work. "Our methods can be run cheaply, adding as little as 1.3% of the original training run’s compute. … The verification strategies we describe are not provably robust, but are intended as an opening proposal" Posed by
| |||
| P089 | Started | Zero-knowledge proofs of training and inferenceWhat is the largest model for which a zero-knowledge proof of training or inference is practical, and at what prover overhead? | 10 |
Zero-knowledge proofs of training and inference. Terms: zero-knowledge proof: a cryptographic proof that a statement is true (e.g. that a computation was done correctly) while revealing nothing else about the inputs. StatusStarted Inference only: a built, open-source ZK proof system measured on LLaMa-2-13B and OPT-13B (one A100 plus 12 CPU cores): 803 s to prove a 2048-token inference, proof under 200 kB, verification in seconds; impractical at production request rates, and later systems (ZKTorch, zkGPT) remain at GPT-2 to Llama-2-7B scale with minutes per token. Training: no ZK proof of full-LLM backpropagation in native floating point has been shown (Peigné-Lefebvre et al. 2026); their 36-month, single-digit-percent-overhead estimate is a paper estimate, and zkLLM's authors write that ZK proofs of LLM training may pose insurmountable challenges. "for LLMs boasting 13 billion parameters, our approach enables the generation of a correctness proof for the entire inference process in under 15 minutes." Posed by
| |||
| P090 | Analyzed | Proof-of-work and proof of declared non-AI useCan a verifier confirm privately, at a small fraction of the job's cost, that a cluster ran a declared non-AI or useful workload? | 2 |
Proof that compute ran a declared non-AI or useful workload. Terms: MatMul(A, B): multiplying matrix A by matrix B. 1 + o(1) overhead: a cost that approaches that of plain matrix multiplication as the matrices grow (o(1) is a term that shrinks toward zero). StatusAnalyzed A proof-of-useful-work protocol for matrix multiplication over a finite field with near-zero prover overhead; security rests on a new hardness conjecture; no implementation, benchmark or attack evaluation. It proves a lower bound on matmul work done, not that a declared workload ran, and says nothing about privacy or about non-AI scientific workloads (weather, CFD, protein folding), which have no work. The Proofworks proof-of-memory protocol (Reuter & Karvonen 2025, preliminary fingerprint experiments) verifies declared AI workloads generally and is a neighbouring compute-accounting problem. "Our main result is a PoUW for the task of Matrix Multiplication MatMul(A, B) of arbitrary matrices with 1 + o(1) multiplicative overhead compared to naïve MatMul. We conjecture that our protocol has optimal security in the sense that a malicious prover cannot obtain any significant advantage over an honest prover." Posed by
| |||
| P091 | Started | Verifiable confidential computing via network commitmentsCan a full loop - the prover's network hardware committing to its traffic, then a neutral site verifying it - give both confidentiality and later proof, demonstrated end to end on a real workload? | 7 |
Architectures where network hardware commits to traffic so a neutral site can later verify confidential workloads. Terms: commitment: a tamper-evident cryptographic 'fingerprint' of data, published so the data cannot later be changed unnoticed. gateway: the network device that records such commitments at the site boundary. StatusStarted One component measured: hashing every frame on one 400GbE link, on a 64-thread server CPU and on a BlueField-3 network card, to build a log for later checking. No commitment protocol or neutral-site verification has been built, and nothing has been demonstrated end to end on a real workload. The design for the full loop is in Harack et al. (2025), which says no existing system does this and that cost, difficulty and time at scale are unanswered. SipHash-1-3-128 clears 400 Gbps at every packet size tested Posed by
| |||
| P092 | Started | Confidentiality and information leakage of verificationIs there a verification design whose channel provably leaks only a declared set of fields, keeping user data and intellectual property private against a state adversary? | 9 |
Information leakage from verification to the verifier, inspector code access, and minimal disclosure. Terms: I(S; Z | Y): conditional mutual information, in bits: how much a released report Z reveals about a protected property S of the workload beyond what the authorised result Y already reveals. StatusStarted Leakage measured as estimated conditional mutual information on real physical-measurement datasets (GPU power, telemetry and electromagnetic traces, including H100 and B200 nodes) across six verification tasks, with one release certified by a Groth16 zk-SNARK; a source reporting its own method. Not a provable information-flow bound: estimation is unreliable for high-dimensional traces, protection holds only under the design distribution of the development data, and nothing is tested against repeated queries, data fusion or a state adversary (Daley: filterability against a state adversary is not established). "Across these tasks, MID produces three releases with perfect held-out verification and zero measured collateral leakage, while the remaining tasks yield explicit privacy–utility frontiers." Posed by
| |||
Workload classification and anomaly detection on telemetry and traffic 6 problems
| ID | Status | Question | Sub-qs |
|---|---|---|---|
| P093 | Started | Workload-type classification from power and telemetryHow accurately can classifiers separate training from inference from non-AI work using telemetry or power, and does that accuracy survive an adaptive evader and hardware drift? | 31 |
Accuracy, evasion robustness and generalisation of classifiers that tell training from inference from non-AI work using on-board telemetry or off-chip/rack/facility power. Terms: β (beta): the largest hidden computation a power trace cannot rule out, as a fraction of the declared machine capacity; β (beta) is its worst case over declared utilisation (how busy the declared work keeps the chip). StatusStarted External power draw of a single NVIDIA H200 NVL sampled at 10 MHz, 5 s traces, 17 LLM families and 25 non-AI workloads, authors' own classifier; four evasion strategies tested by the authors, who say this is not a comprehensive adversarial evaluation: three caught at or above 99%, diluted LoRA at 48-88% without the rescue rule, below this survey's detection goal of 90%; cross-device and cross-generation generalisation 'remains to be established'; single node. Corroborating own-method report: Rahman & Tajdari (2026) classify training against non-training from zero-overhead GPU telemetry at 98.2% across 9 GPU types, single node up to 8 GPUs, with dilution and idle-padding evasion tested. Harack et al. (2025) 'not yet possible to reliably know' predates both. "we separate training from inference and from non-AI computation with an accuracy of 97% and a macro-averaged F1 score of 0.955, evaluated on model families unseen during training" … "The fourth, diluted low-rank adaptation (LoRA), is detected 48–88% of the time with a hardened classifier" Posed by
| |||
| P094 | Posed | Purpose-level (semantic) use verificationCan any on-chip-measurable feature verify the purpose or beneficiary of a workload, not just its type? | 1 |
Verifying the purpose or beneficiary of a workload, not just its type. StatusPosed Posed by
| |||
| P095 | Started | Hidden concurrent compute on declared hardwareDoes a hidden concurrent workload produce a measurable signal, and how much compute can be freed undetectably on declared hardware? | 2 |
Hidden concurrent compute and memory on declared hardware. StatusStarted Hidden inference on parallel CUDA streams alongside re-executed LLM inference on identical hardware, detected by timing with numerical fingerprinting as a complement; the author's own experiments, unquantified (no fraction of declared compute, no detection rate). The second clause, how much compute can be freed undetectably, has only paper-only work: Reuter's expectation for his own proposed protocol (1-2%) and a verifiable-memory-exhaustion draft with placeholder results. Bounding hidden memory and model-swap time: nothing built. Any significant hidden workload would contend for resources, causing detectable slowdowns. In my experiments, resource contention was evident in timing. Posed by
| |||
| P096 | Analyzed | Inference deployment monitoringCan an operator prove all inference passed through a given monitor, and how well can misuse or diversion be caught among approved traffic? | 9 |
Monitoring inference deployments for coverage, misuse and diversion. StatusAnalyzed A design: a verifier-developed classifier that the deployer must run on all inputs and outputs, with coverage ensured by weight containment plus cluster-level batch authorization; the authors rate feasibility 'Medium; 1-2 years'; nothing built or measured for either clause. Proving coverage to an outside inspector: paper only (also flexHEG Part 1's inference receipts). Catching misuse among approved traffic at a stated false-positive rate: paper only; the authors argue low sensitivity may be tolerable. Compute diverted before telemetry flags it: Hausenloy & Li's second-hand claim is not evidence. One major hurdle is ensuring that all relevant inputs and outputs are passed through the classifier. One way to achieve this is to both: prevent weights from leaving the data center, and have close monitoring of all chips in the data center. Posed by
| |||
| P097 | Analyzed | Detecting distributed training from network trafficCan a classifier on network-traffic metadata detect low-communication distributed training at high true-positive and low false-positive rates, and does it survive deliberate traffic shaping? | 4 |
Detecting distributed training from network traffic. StatusAnalyzed A desk analysis of the headline's question: detection must use volume and routing; DiLoCo-style training at under 40 Mbps sits below average household upload speed; jitter, variable-rate streaming, pauses and VPN routing can disguise it, so the author expects the requirements to favour evaders. The 40 Mbps figure comes from the paper's simulation, not a measured run. No classifier designed, trained or tested on traffic; no detection or false-positive rate for this survey's goals. Moon et al.'s large-data-transfer threshold is a proposal for detecting training through cloud ingress, a neighbouring question. Koopmanschap & Barten's note that concealment removes the signature is an argument. any efforts at detecting distributed training by monitoring internet traffic would have to rely on analysis of the volume and routing, not on the content … the traffic can be disguised into almost any irregular pattern. Traffic shaping methods include jitter, variable-rate streaming, and intermittent pauses. Posed by
| |||
| P098 | Started | AI-enabled verificationWhen will AI auditors be competent and secure enough to deploy, and can they resist jailbreaks, prompt injection and backdoors by a sophisticated adversary? | 0 |
AI-enabled verification and auditor robustness. StatusStarted A built LLM-auditor prototype (TEE and attestation emulated on a CPU without SEV; frontier model called remotely) evaluated on 21 real codebases: perfect discrimination of correct from wrong papers on objective questions, less stable on subjective ones; the audit task is software-artifact verification, not compute verification. The 'when' clause has no timeline estimate. Resistance to jailbreaks, prompt injection and backdoors: not tested; the author concedes prompt injection cannot be prevented and designs for it to leave evidence in a signed transcript. We demonstrate this architecture by automating the artifact evaluation process for 21 peer-reviewed computer science papers with released codebases on GitHub (e.g. Does the codebase implement the system described in the paper?). We verified five high-level properties of these codebases described in the corresponding publications, treating the source code as private. Posed by
| |||
Distributed training versus bandwidth, run lineage and model provenance 7 problems
| ID | Status | Question | Sub-qs |
|---|---|---|---|
| P099 | Analyzed | Hardware scope and chip classificationWhich chip specifications predict training throughput well enough to set a regime's hardware scope, and where do consumer/edge parts fall relative to that line? | 8 |
Which chips and specs a regime covers, how chip capability is classified and measured, and where consumer-GPU and edge cases fall. Terms: H100-equivalent (H100e): a common unit that counts mixed accelerators as multiples of one Nvidia H100 data-centre GPU, so fleets of different chips can be compared on one scale. N: the factor by which a cluster is larger when each chip has 1/N of the throughput. StatusAnalyzed Simulated (a distributed-training cost model calibrated to Covenant-72B, not a measured run): adds a 1,280 GB HBM memory threshold beside the 16-H100-equivalent cluster line for a 10^25 FLOP evasion, showing memory capacity, not only FLOP/s, matters for scope. Consumer/edge clause: the same desk analysis places 4-8 GPU workstations, DGX-class servers and GDDR parts below that threshold, also paper only. No throughput-prediction study; nothing measured. '1025' is the saved text's rendering of 10^25. imposing the memory limit reduces the maximum unregistered per-node model from 240B to roughly 91B parameters, forces the evader to operate approximately five times as many nodes for equivalent model quality, and raises the cost of illicitly developing a 1025 FLOP model by roughly 50%. Posed by
| |||
| P100 | Analyzed | Compute-threshold definitions and loopholesWhich compute counts toward a governance threshold, how are thresholds set and renegotiated, and how can runs be structured to evade them? | 4 |
What compute counts toward thresholds, how thresholds are set, moved and renegotiated (including regional caps and coverage exemptions), and how runs evade them. StatusAnalyzed Evasion clause only: a simulated cost of structuring a 10^24, 10^25 or 10^26 FLOP run as clusters below every proposed registration threshold (DiLoCo-family methods over 100 Mbps links; extrapolated 10-300x beyond tested model size). The same paper proposes a 1,280 GB HBM memory threshold as a fix to the threshold definition, also paper only. Which compute counts (fine-tuning, distillation, mixture-of-experts, inference scaling) and how thresholds are renegotiated: no work beyond framing and stated difficulties. these proposed or existing regulations could be evaded using $1.6M, $31M, and $3.8B, respectively, of hardware arranged in clusters smaller than any proposed registration requirement Posed by
| |||
| P101 | Started | Run lineage, splitting and launderingCan split, resumed or laundered training runs be linked to one model, at a stated false-link rate? | 4 |
Linking split, resumed or laundered runs to one model. StatusStarted Experiments on GPT-2 (124M parameters, OpenWebText) and published Pythia checkpoints: a glued transcript (two training runs reported as one continuous run, i.e. a run resumed from another run's checkpoint) is caught by a weight-norm jump and a memorization cut-off at the join; the verifier holds data, weights, code and checkpoints. The opposite direction (one long run reported as many 'fresh' short runs) rests on the certified-random initialisation design, argued hard to spoof but not tested; no false-link rate at frontier scale; laundering of a checkpoint through declared 'training data' not covered. Below, we show that our methods address existing attacks from the literature (Glue-ing and Interpolation), and also demonstrate our method’s response to two new attacks (Data Addition and Subtraction). Posed by
| |||
| P102 | Analyzed | Low-communication distributed training versus bandwidthAt what inter-site/inter-pod bandwidth does low-communication distributed training become uneconomic relative to a co-located baseline for a frontier run? | 18 |
Efficiency of multi-site / low-communication training as a function of inter-site or inter-pod bandwidth, and the bandwidth thresholds that would make it uneconomic. Terms: low-communication distributed training (e.g. DiLoCo): training methods that let machines in different places cooperate while exchanging far less data than usual, cutting the bandwidth needed between sites. n: GPUs per pod. B: the cap on external link bandwidth. T: the time allowed for training. StatusAnalyzed Simulated cost for 10^25 FLOP of local-equivalent compute with DiLoCo-family methods over 100 Mbps links and nodes below 16 H100-equivalents; extrapolated 10-300x beyond tested model size and up to 250x in replica count; calibrated against Covenant-72B but not a measured run. No bandwidth at which training becomes uneconomic is identified; the paper's cost table spans 10 Mbps to 1 Gbps. At 100 Mbps bandwidth, it costs about 3× more to do training distributed relative to centralized, but it is not a substantial barrier to feasibility. Posed by
| |||
| P103 | Started | Training-data verificationCan a verifier confirm a declared dataset's provenance and composition produced a given model, and detect a tiny fraction of backdoor data? | 4 |
Verifying training data provenance and composition. StatusStarted Proof-of-Training-Data tested on GPT-2 (124M) and Pythia (70M) checkpoints: catches gluing, interpolation, data subtraction and large data additions for about 1.3% of training compute; the verifier is given data, weights and code. The headline's backdoor clause is not met: the authors concede the method is not robust to small additions of harmful data (quoted by Reuel et al.). Composition rules (e.g. code-token fraction) and privacy-preserving membership proofs have designs only (Scher 2026; flexHEG Part 1); verifying that claimed data origins are uncompromised: no work. We show experimentally that our verification procedures can catch a wide variety of attacks, including all known attacks from the Proof-of-Learning literature. Posed by
| |||
| P104 | Started | Training-procedure and code verificationCan a declared architecture, optimiser schedule and code be checked against an executed run, catching a deliberately mis-declared one? | 4 |
Checking training procedure and code against declarations. StatusStarted Delegation setting: a client fixes the program and several untrusted providers run it; a referee localises any divergence to one training step and one operator, re-executes it and rules, so an execution that departs from the declared program is caught given bitwise reproducibility (the RepOps library). Implemented and benchmarked on DistilBERT and Llama-3.1 1B and 8B training and inference, FP32 only, about 60% overhead on an A100. Not shown: a prover who both writes and mis-declares the code, frontier scale, or any check of the declaration's content (architecture, optimiser schedule) beyond re-execution; Choi et al. (2023) re-execute segments of a declared transcript on GPT-2 in the same spirit, for data rather than code. Our implementation shows that refereed delegation achieves both strong guarantees for clients and practical overheads for compute providers. Posed by
| |||
| P105 | Started | Model identity across evaluation and deploymentCan a verifier confirm the evaluated, deployed and served models are the same (e.g. via attested weight/code hash) in a real pilot? | 3 |
Showing evaluated, deployed and served models are the same. StatusStarted Served-model identity by recomputation on one 8xH100 server with prover and verifier on the same box; open-weight 7B-120B models; developer-run attacks (model substitution, LoRA append, append-on-trigger); 37,500 prompt runs; a fixed threshold gave honest-prover failures and short-output false negatives. The evaluated-versus-deployed clause (a TEE-signed hash of weights and code) is a design only, which Scher & Thiergart call 'mostly solved in principle'; no real pilot, which the headline asks for; handling of legitimate model versions: no work. Our results add further evidence that Token DiFR can identify different incorrect-model attacks reliably across a small range of prompts. Posed by
| |||
Inspection practice and verification institutions with a technical core 9 problems
| ID | Status | Question | Sub-qs |
|---|---|---|---|
| P106 | Analyzed | On-site physical inspection capacity and practiceHow many inspector-days, at what cost and sample size, are needed to verify a given number of accelerators or data centres at a stated detection probability? | 10 |
Inspector-hours, sample sizes, cost and procedures for physical inspection of chips, racks and data centres. Terms: N: the number in question, of chips in a hidden cluster, or of chips, racks, accelerators or data centres to be inspected. StatusAnalyzed A design and estimate from 2023: random sampling of chips' logged weight snapshots, with a formula for the samples needed to catch a rule-breaking run at a given probability. For a 10^25 FLOP run among 10^7 chips, caught within 30 days at 90% probability: about 232,000 samples a year, or about 232 inspectors if each checks 1,000 chips a year (an assumption), against 280 IAEA inspectors in 2021. No cost estimate and nothing run; it assumes chips that log weight snapshots, which still do not exist in 2026, and no later paper has updated the numbers. The installed base has roughly reached his assumption (over 15 million H100-equivalents in January 2026, Epoch AI), and 10^25 FLOP is now an ordinary run rather than a frontier one. This would require ≈ 232 inspectors, slightly smaller than the 280 active IAEA inspectors as of 2021. Posed by
| |||
| P107 | Posed | Inspection notice and concealment timeHow quickly can accelerators, logs and storage be moved, hidden or wiped, and so how short must inspection notice be? | 1 |
How quickly accelerators, logs and storage can be moved, hidden or wiped, and so how short inspection notice must be; including highest-assurance audits, distributed training networks and keeping protocols current. StatusPosed Posed by
| |||
| P108 | Analyzed | Inspection triggers and the will to invoke themWhat should trigger an inspection, and how can inspections still happen when the suspect state has a great-power patron? | 1 |
What triggers an inspection, and whether inspections happen at all when the suspect state has a great-power patron or the host state denies cooperation. StatusAnalyzed A proposed design (automatic compute-anomaly triggers inside a three-layer inspectorate) that answers the trigger clause on paper; the great-power-patron clause rests on the authors' reading of the historical record, and their analysis of challenge inspections in the Alva Myrdal Centre treaty database is not yet reported; no false-alarm rate and no trigger calibration, which the authors call 'the substantive work ahead'. The Proofworks observation that no challenge inspections have been implemented between great powers and Koopmanschap & Barten's point that power monitoring needs utility cooperation are context and a stated limit. formal challenge inspections historically go uninvoked when a violator has a great-power patron, so verification cannot depend on a state being willing to accuse. Raised in: Koopmanschap & Barten (2026), How to Catch a GPU: A Taxonomy of Verification and Enforcement Mechanisms for International AI Agreements; Leklem & Maheshwari (2026), How Should We 'Inspect' Artificial Intelligence? (AMC Insights); Proofworks, Living Verification Documents. | |||
| P109 | Posed | Managed access and disclosure limitsUnder managed access, what categories of sensitive information will provers refuse to disclose, and can inspections still give the required guarantees? | 2 |
Managed access and limits on what provers disclose. StatusPosed Posed by
| |||
| P110 | Analyzed | Personnel-based verification and whistleblowersWhat certainty can personnel-based verification and whistleblower channels give, and can they survive a prover that compartmentalises, surveils or intimidates staff? | 2 |
Personnel-based verification and whistleblower channels. StatusAnalyzed Whistleblower clause only: a draft treaty article (Article X) with anti-retaliation duties, confidential or anonymous reporting channels run by the treaty body, identity protection, and asylum for whistleblowers and their families; a proposal, untested. The certainty personnel-based verification can give: no work beyond Harack's stated limit ('achievable certainty might be much lower than ... with hardware'). Survival against staff under video surveillance or with backdoored computers: Baker et al. state the limit, no design. Reach into state-run operations abroad: Rahman states the limit. Parties prohibit and prevent retaliation against Covered Whistleblowers and Associated Persons, including but not limited to dismissal, demotion, blacklisting, loss of benefits, harassment, intimidation, threats, civil or criminal actions, visa cancellation, physical violence, imprisonment, restriction of movement, or other adverse measures. Posed by
| |||
| P111 | Analyzed | Escalation and attribution of findingsWhat escalation ladder, graded-evidence aggregation and attribution procedure resolve ambiguous findings and assign a violation to the responsible party? | 1 |
Escalation ladders for ambiguous findings, aggregation of graded evidence, and attribution of a violation to the responsible party. StatusAnalyzed A draft treaty article (Article XI, modelled on Articles IX, XII and XIV of the Chemical Weapons Convention): notification, 36-hour acknowledgement and 5-day clarification, Executive Council adjudication with a challenge inspection, then graded measures (added monitoring, hardware relinquishment, sanctions, protective actions); a proposal, untested. Aggregating graded evidence (many sub-threshold acts) and attributing a violation to the responsible actor (cloud provider versus customer, runs split across providers) have no design; Moon et al.'s call for an agency able to unmask evaders is framing. The Concerned Party notifies the Requested Party of their concern, while also sharing their concern with the Director-General and Executive Council. The Requested Party will acknowledge this notification within 36 hours, and provide clarification within 5 days. Raised in: Baker, Kulp, Marks, Brundage & Heim (2025), Verifying International Agreements on AI: Six Layers of Verification for Rules on Large-Scale AI Development and Deployment (RAND WR-A4077-1); Scher & Thiergart (2024), Mechanisms to Verify International Agreements About AI Development (MIRI); Aarne & Petrie (2025), International Security Applications of Flexible Hardware-Enabled Guarantees (flexHEG Part 3); Cankaya (2026), Compute Verification FAQs; Ansari (2026), Hardware-Level Governance of AI Compute: A Feasibility Taxonomy for Regulatory Compliance and Treaty Verification; Daley (2026), Proving the Negative: Three Warnings from Arms Control for the Age of AI (War on the Rocks); Moon, Vedula, Geneson & Bar-on (2025), Strategies and Detection Gaps in a Game-Theoretic Model of Compute Governance (RAND RRA3686-1). | |||
| P112 | Started | Replay mismatches and dispute settlementWhat is the baseline rate of benign replay mismatches, and can a dispute procedure attribute a mismatch to hardware fault versus cheating? | 4 |
The rate of benign mismatches in replay checks, and whether a dispute procedure can tell a hardware fault from cheating. StatusStarted Delegation setting: when two executions of the same ML program disagree, a referee bisects by checkpoint hashes to one training step and one operator, re-executes it and rules; benign floating-point mismatches are removed by bitwise-reproducible operators, implemented and benchmarked on DistilBERT and Llama-3.1 1B and 8B in FP32. It finds which party is wrong, not whether the cause was a hardware fault or cheating; no baseline rate of benign replay mismatches is measured; Cankaya (2026) claims his own bitwise replay settles statistical false alarms, a self-report at the same stage; GPU-stack training replay at frontier scale not shown. we design Verde, a dispute arbitration protocol that efficiently handles the large scale and graph-based computational model of modern ML programs. For (2), we build RepOps (Reproducible Operators), a library that eliminates hardware “nondeterminism” by controlling the order of floating point operations performed on all hardware. Our implementation shows that refereed delegation achieves both strong guarantees for clients and practical overheads for compute providers. Raised in: Cankaya (2026), A System Overview for Near-Term, Low-Trust AI Compute Verification (v0.2); Cankaya (2026), Catching Misreporting About ML Hardware Use by Turning Noise into Signal, Part II; Cankaya (2026), Compute Verification FAQs; Canary Institute research. | |||
| P113 | Posed | Verification facilities and labsWhat jointly operated verification facilities/labs exist or have been costed, and what equipment, staffing and lead time does each verification task need? | 4 |
Verification facilities, labs and their capacity. StatusPosed Posed by
| |||
| P114 | Started | End-to-end pilots, demonstrations and red-team exercisesWhat have published pilots and red-team exercises of the verification stack (including two-state experiments) detected, leaked, and how fast did they warn? | 8 |
Running and publishing prototypes, pilots and red-team events of the verification stack, including two-state experiments. StatusStarted One 8xH100 server with prover and verifier on the same box; open-weight 7B-120B models; inference only; developer-run attacks (model substitution, LoRA append, append-on-trigger); 37,500 prompt runs; completeness ignored; a fixed threshold gave honest-prover failures and short-output false negatives. Not an independent red team, no national-security team, no two-state experiment, no leak or warning-time measurement; Reynolds & Foster's line-rate hashing benchmark is a second self-run component test. Red-team exercises of the stack and a joint international experiment: no work (Baker et al., early 2025: the listed challenges 'mostly receive little to no work'). Our results add further evidence that Token DiFR can identify different incorrect-model attacks reliably across a small range of prompts. Posed by
| |||
Institutional and framing questions 13 problems
| ID | Status | Question | Sub-qs |
|---|---|---|---|
| P115 | Analyzed | Enforcement and responseWhat response follows detection, how is it made proportionate and deterrent, and how is enforcement separated from verification? | 0 |
Response and enforcement after detection. StatusAnalyzed A draft treaty's proposed response ladder (diplomacy, sanctions, escalation to disruption), with the same text separating detection of covert violations from nonproliferation and enforcement; proportionality is specified only for domestic penalties on restricted research (Article IX: 'proportionate to the severity of the violation and ... designed to act as a sufficient deterrent'). Nothing modelled, tested or exercised; no attribution procedure. Baker et al. (RAND 2025, Table 4) list the same options (reciprocation, sanctions, export controls, withdrawal of guarantees, cyber or kinetic attacks, technical mechanisms) and note they 'would ideally be effective deterrents and thus never carried out'; Harack et al. put most enforcement outside their scope. "Should nonproliferation fail, or if a rogue actor is found to be working towards ASI, coalition members can employ their agreed-upon enforcement mechanisms. These start with standard tools of international pressure such as diplomacy and economic sanctions. Depending on the nature of the developments, countries could escalate to other means at their disposal to disrupt, slow, or impede ASI development." Posed by
| |||
| P116 | Started | International verification body: powers and governanceWhat powers, governance, staffing and access should an international AI verification body have? | 2 |
The powers, governance, staffing, inspector model, access and registry governance of an international AI verification body. StatusStarted An empirical result from the Alva Myrdal Centre Arms Control Agreement Database (99 compliance mechanisms, 33 treaties, 1959 to 2015) on one leg only, the inspector model. Powers, decision-making, staffing, trigger calibration, access provisions and registry governance have paper proposals at most (the authors' three-layer inspectorate; Harack et al. and Scher treat body design as political and out of scope), and the authors call them 'the substantive work ahead'. The scope of the agreement is strongly associated with the inspector model. Bilateral treaties use party-on-party inspectors 95% of the time; multilateral treaties use institutional inspectors 47.4% of the time, roughly 10× the bilateral rate of 4.9%. Raised in: Harack et al. (2025), Verification for International AI Governance (Oxford AIGI); Baker, Kulp, Marks, Brundage & Heim (2025), Verifying International Agreements on AI: Six Layers of Verification for Rules on Large-Scale AI Development and Deployment (RAND WR-A4077-1); Shavit (2023), What Does It Take to Catch a Chinchilla? Verifying Rules on Large-Scale Neural Network Training via Compute Monitoring; Scher (2026), Verifying Restrictions on Frontier AI Research (MIRI); Ilhan, Withers, Gietz & Harack (2026), Verifiable Semiconductor Manufacturing (Oxford AIGI); Hausenloy & Li (2026), The Compute Verification Post; Koopmanschap & Barten (2026), How to Catch a GPU: A Taxonomy of Verification and Enforcement Mechanisms for International AI Agreements; Leklem & Maheshwari (2026), How Should We 'Inspect' Artificial Intelligence? (AMC Insights); Wasil, Reed, Miller & Barnett (2024), Verification Methods for International AI Agreements; AI Security Forum, Requests for Discussion. | |||
| P117 | Posed | Security and legibility of the verification bodyCan a verification body run its own operations securely against infiltration, and be transparent enough to be trusted without leaking what it holds? | 1 |
Whether a verification body can run its own operations securely against infiltration and be transparent enough to be trusted without leaking what it holds. StatusPosed Raised in: Harack et al. (2025), Verification for International AI Governance (Oxford AIGI); Baker, Kulp, Marks, Brundage & Heim (2025), Verifying International Agreements on AI: Six Layers of Verification for Rules on Large-Scale AI Development and Deployment (RAND WR-A4077-1); Ilhan, Withers, Gietz & Harack (2026), Verifiable Semiconductor Manufacturing (Oxford AIGI); Rahman & Tajdari (2026), Detecting Hidden ML Training With Zero-Overhead Telemetry (MIRI). | |||
| P118 | Analyzed | Translating treaty commitments into checkable rulesWhich arms-control commitments can be translated into technically checkable rules, and how? | 0 |
Which arms-control commitments can be turned into technically checkable rules, and how machine-checkable artifacts map to treaty commitments. StatusAnalyzed A paper mapping of candidate agreement types (compute caps, mandated evaluations, public reporting, non-proliferation, ruleset-based agreements) onto claims a flexHEG device could check; the same report poses the translation question as open for military cyber-physical systems. Harack et al. (2025) analyse verification requirements by agreement type in their section on agreements. The mapping from machine-checkable artifacts (receipts, audit trails) to legal instruments raised by AI 2040 has no worked example; nothing implemented. Verification of such claims could be used to verify compliance with formal agreements regarding any of the above, such as agreements not to train models above a given size or capability level, agreements to ensure that models always need to attain sufficient scores on particular safety evaluations, or agreements to simply publicly report the results of verifiable evaluations. Posed by
| |||
| P119 | Started | Legal limits on verification operationsWhat legal limits apply to verification operations: retaining captured data across borders, defensive jamming, and non-interference with national technical means? | 2 |
The legal limits on verification operations: retaining captured data across borders, defensive jamming, sensors in territorial waters, and non-interference with national technical means. StatusStarted One leg only, non-interference with national technical means: a count from the Alva Myrdal Centre treaty database (99 mechanisms, 33 treaties, 1959 to 2015) showing such clauses exist only in bilateral treaties, so a multilateral AI treaty's clause would be novel; Scher (2026) notes SALT I to New START authorise NTM and prohibit interference (precedent). Cross-border retention of captured data (Cankaya: 'will vary on a case-by-case basis') and defensive jamming (Cankaya: 'out-of-scope') have no work. Teague et al.'s estimate that passive acoustic sensors would need to sit inside the 12-nautical-mile territorial sea is for the statement's territorial-waters clause, not the headline (related). All 10 NTM non-interference clauses and all 7 concealment prohibitions in the dataset are bilateral. Raised in: Scher (2026), Verifying Restrictions on Frontier AI Research (MIRI); Cankaya (2026), Research Note: The Fundamentals and Feasibility of Secure Network Taps for Verifying AI Datacenter Use; Cankaya (2026), Suppressing Side Channels in an Untrusted Data Center via Retrofitted Defenses (MIRI); Koopmanschap & Barten (2026), How to Catch a GPU: A Taxonomy of Verification and Enforcement Mechanisms for International AI Agreements; Leklem & Maheshwari (2026), How Should We 'Inspect' Artificial Intelligence? (AMC Insights); Teague, Rajmohan & Muehlhaeuser (2026), Could Underwater Data Centers Pose a Risk to AI Treaty Verification?, arXiv 2609.18824. | |||
| P120 | Analyzed | Participation and coverage across statesHow can declarations and coverage be extended across the compute-hosting states, and how do non-participants' capabilities affect a partial agreement? | 0 |
Participation and coverage across states. StatusAnalyzed A plan-level proposal: US–China mutual declaration first (about 85% of global compute), then extension to the 5–10 other compute-hosting nations to reach 99%, with urgency tied to intelligence uncertainty about those nations' compute; no mechanism for securing participation and no modelling of a partial agreement. Baker et al. (RAND 2025, Table 4) list options for attaining participation (export controls and energy policy to keep hardware in cooperative states, dialogues, leverage). Non-participants clause: Harack et al. (2025) note only that protocols for non-global agreements 'may have to take into account the existence of non-participating states'; Daley poses the dependence on other states; host-state seizure (Harack et al.) is a stated risk. "Extend this declaration to other countries with significant compute ownership or production capacity. Do so urgently unless there is confident intelligence in place that would detect this compute being moved." Posed by
| |||
| P121 | Analyzed | Building trust in shared mechanism designsDoes shared design documentation plus adversarial review build enough international trust in shared mechanisms, or will states insist on de novo designs? | 3 |
Building international trust in shared mechanism designs and standards. StatusAnalyzed A qualitative desk comparison of three routes to a trusted standard (adapting existing commercial designs, open-source projects such as OpenTitan, seL4 and Caliptra, and a de novo international collaboration), weighing the security limits of commercial designs and the precedent of weakened encryption standards; no review conducted, no standard written, no state's position recorded. Standards for assessing other parties' solutions, trust in lab findings and an appliance standard: calls for work only. "Concerns about subtle vulnerabilities could likely be addressed through a detailed international review of the design, where each design choice is carefully justified to ensure that it was not animated by any ulterior motive. However, sufficiently suspicious states may still insist that the only acceptable long-term solution is a de novo design." Posed by
| |||
| P122 | Analyzed | Field capacity, funding and deployment readinessWhat does the verification field need, in funding, people and manufacturing capacity, to reach deployment, and how does that compare with what exists? | 2 |
Field funding, capacity and deployment readiness. StatusAnalyzed Funding only, as a proponent's unsourced figures: a $50 to 150M capitalisation recommended for one central project, against 'less than $15M' spent on verification technologies to date (same document, AI-lab section). People and manufacturing capacity are not estimated: the companion SITREP lists 'preparing plans for quick and secure scaled-up manufacturing' as an extension and Proofworks notes a multi-month manufacturing buffer. No survey of the field's funding or headcount. Fund a large central project to “own” Plan A verification. Their goal should be to ensure AI verification technology can be deployed in early 2029. The remit would cover not just the R&D but also the stockpiling and planning for scale-out. Before the end of 2026, they should be capitalised with at least $50-150M, and as the project succeeds, they should expect to scale to absorb more funding. Raised in: AI Security Forum, Requests for Discussion; Amodo (2026), AI 2040 Plan A — How You Can Help; Amodo (2026), AI 2040 Plan A — Verification SITREP; Proofworks, Living Verification Documents. | |||
| P123 | Posed | Choice of technical approachDoes any approach beat network tap + recomputation on cost and assurance, and should verification lean more on personnel/intelligence methods? | 0 |
Whether current technical bets (tap + recomputation) are right. StatusPosed Posed by
| |||
| P124 | Posed | Validity of compute as the governed quantityDoes frontier capability still require large-scale training, and how fast does algorithmic progress make a fixed compute threshold obsolete? | 4 |
Whether compute remains a valid proxy for capability. Terms: N: model size. M: total communication between pods. StatusPosed Posed by
| |||
| P125 | Analyzed | What claims verification should targetWhich verification claims are both useful and feasible — e.g. "civilian not military", "not building prohibited AI", a speed limit — and at what evidence granularity? | 0 |
Which claims verification should target. StatusAnalyzed A desk mapping of agreement families (transfer knowledge, transfer resources, pool resources, prepare for emergencies, regulate development and deployment) to verification requirements and to mechanisms judged effective, mature and politically viable, including the conclusion that 'a true ban is likely impossible via blunt mechanisms, but is potentially workable via fine-grained controls'. The 'civilian not military' and 'not building prohibited AI' claims are posed by Harack (2026) without a feasibility result; a speed limit or pause: no work. Evidence-granularity clause: Cankaya asserts collision-resistant per-workload capture as a requirement and asks whether token-level capture is reasonable (self-deprecated v0.2); no work beyond that. Nothing piloted or negotiated. "This section provides an overview of possible international agreements relating to AI and their verification requirements." … "These were chosen because they appear to be a compelling combination of a) effective, b) technologically mature, and c) politically viable." Posed by
| |||
| P126 | Analyzed | Independent readiness reviewHave the mechanisms that verification proposals depend on had an independent technology-readiness review, and do independent ratings match the proponents' own? | 7 |
Whether the mechanisms verification proposals depend on have had an independent technology-readiness review, whether independent ratings match the proponents' own, and the deployment timelines the sources give per mechanism family. StatusAnalyzed A literature-based feasibility taxonomy of about twenty hardware-governance mechanisms by an author outside the proponent groups; by its own limitations section the ratings come from the author's reading 'rather than from formal expert elicitation or independent technology readiness review'; no comparison of independent ratings against proponents' own. The per-family timelines in the sources (Cankaya: off-chip mechanisms in testing and de-risking as of July 2026; Harack et al.: a one-year crash program; Aarne et al.: about four years for on-chip uptake) are proponents' estimates and belong to those families' problems. Our analysis reveals a structural mismatch: the mechanisms most needed for treaty verification … are also the least mature. Raised in: Harack et al. (2025), Verification for International AI Governance (Oxford AIGI); Petrie, Aarne, Ammann & Dalrymple (2025), Overview of Flexible Hardware-Enabled Guarantees (flexHEG Part 1); Petrie & Aarne (2025), Technical Options for Flexible Hardware-Enabled Guarantees (flexHEG Part 2); Aarne, Fist & Withers (2024), Secure, Governable Chips (CNAS); Scher (2026), Verifying Restrictions on Frontier AI Research (MIRI); Cankaya (2026), All Hands on Deck to Build the Datacenter Lie Detector; Cankaya (2026), Compute Verification FAQs; Canary Institute research; Ansari (2026), Hardware-Level Governance of AI Compute: A Feasibility Taxonomy for Regulatory Compliance and Treaty Verification; Harack (2026), AI Verification: Infrastructure for Prosperity, Governance, and Peace (Lawfare); Koopmanschap & Barten (2026), How to Catch a GPU: A Taxonomy of Verification and Enforcement Mechanisms for International AI Agreements; AI 2040 (2026), Get Involved in Verification; Dean, Verification Plan (AI 2040); Proofworks, Living Verification Documents. | |||
| P127 | Started | Cost and overheadWhat total overhead (as % of training/facility cost) does the full verification stack impose in a pilot, and can it stay far below the monitored hardware's cost? | 4 |
Total cost and overhead of verification. StatusStarted One mechanism measured in the lab, not a stack in a pilot: making a small Llama model's computation reproducible for dispute checking costs about 60% extra time for training and inference on one A100. Device-cost estimate: monitoring hardware of hundreds to a few thousand dollars per rack against racks costing millions (Cankaya 2026, an order-of-magnitude estimate that leaves out guarding, inspection, recomputation and staff). No full verification stack has been piloted, so a cost of about 1% of a facility's cost is not shown. both inference and training on Llama-3.11B incur around 60% overhead RepOps on an A100 GPU Posed by
| |||
Documents
- Aarne & Petrie (2025), International Security Applications of Flexible Hardware-Enabled Guarantees (flexHEG Part 3)
- Aarne, Fist & Withers (2024), Secure, Governable Chips (CNAS)
- Abdelghafar & Kulp (2026), Privacy-Preserving AI Verification via Minimal Information Disclosure
- AI 2040 (2026), Get Involved in Verification
- AI Security Forum, Requests for Discussion
- Alpay & Alpay (2026), Unprivileged Topology Certificates for Cloud GPU Attestation, arXiv 2606.24934
- Amodo (2026), AI 2040 Plan A — How You Can Help
- Amodo (2026), AI 2040 Plan A — Verification SITREP
- Amodo (2026), An Inference Verification Prototype — Stage 1
- Amodo (2026), Example Schemes for Verifying High-Stakes AI Agreements
- Ansari (2026), Hardware-Level Governance of AI Compute: A Feasibility Taxonomy for Regulatory Compliance and Treaty Verification
- Baker, Kulp, Marks, Brundage & Heim (2025), Verifying International Agreements on AI: Six Layers of Verification for Rules on Large-Scale AI Development and Deployment (RAND WR-A4077-1)
- Canary Institute research
- Cankaya (2025), Catching Misreporting About ML Hardware Use by Turning Noise into Signal
- Cankaya (2025), TSMC Most Definitely Has a Golden Record of All AI Chips It Made
- Cankaya (2026), A System Overview for Near-Term, Low-Trust AI Compute Verification (v0.2)
- Cankaya (2026), All Hands on Deck to Build the Datacenter Lie Detector
- Cankaya (2026), Bit-Exact AI Inference Verification Without Performance Tradeoffs, arXiv 2606.00279v2
- Cankaya (2026), Catching Misreporting About ML Hardware Use by Turning Noise into Signal, Part II
- Cankaya (2026), Compute Verification FAQs
- Cankaya (2026), Research Note: The Fundamentals and Feasibility of Secure Network Taps for Verifying AI Datacenter Use
- Cankaya (2026), Suppressing Side Channels in an Untrusted Data Center via Retrofitted Defenses (MIRI)
- Daley (2026), Proving the Negative: Three Warnings from Arms Control for the Age of AI (War on the Rocks)
- Dean, Verification Plan (AI 2040)
- Gargiulo & Kulp (2026), Workload Identification with Physical Side Channels for AI Governance
- Harack (2026), AI Verification: Infrastructure for Prosperity, Governance, and Peace (Lawfare)
- Harack et al. (2025), Verification for International AI Governance (Oxford AIGI)
- Hausenloy & Li (2026), The Compute Verification Post
- Ilhan, Withers, Gietz & Harack (2026), Verifiable Semiconductor Manufacturing (Oxford AIGI)
- Jin et al. (2026), Trusting What You Cannot See: Auditable Fine-Tuning and Inference for Proprietary AI, arXiv 2603.07466
- Karvonen et al. (2025), DiFR: Inference Verification Despite Nondeterminism
- Kimpson, Baker & Graham (2026), Can Power Draw Constrain Covert Compute? Limits of Analogue Verification for AI Governance, arXiv 2610.07476
- Koopmanschap & Barten (2026), How to Catch a GPU: A Taxonomy of Verification and Enforcement Mechanisms for International AI Agreements
- Kulp et al. (2024), Hardware-Enabled Governance Mechanisms (RAND WR-A3056-1)
- Leklem & Maheshwari (2026), How Should We 'Inspect' Artificial Intelligence? (AMC Insights)
- Milton & Reynolds (2026), Network Tapping for AI Verification: A Technical Assessment
- Moon, Vedula, Geneson & Bar-on (2025), Strategies and Detection Gaps in a Game-Theoretic Model of Compute Governance (RAND RRA3686-1)
- Peigné-Lefebvre, Nguyen & Wang (2026), Zero Knowledge Verification for Frontier AI Training Is Possible, arXiv 2606.05433
- Petrie & Aarne (2025), Technical Options for Flexible Hardware-Enabled Guarantees (flexHEG Part 2)
- Petrie, Aarne, Ammann & Dalrymple (2025), Overview of Flexible Hardware-Enabled Guarantees (flexHEG Part 1)
- Proofworks, Living Verification Documents
- Rahman & Tajdari (2026), Detecting Hidden ML Training With Zero-Overhead Telemetry (MIRI)
- Rahman (2026), Does Distributed Training Undermine Compute Governance? (MIRI)
- Reuel, Bucknall et al. (2025), Open Problems in Technical AI Governance, compute sections
- Reynolds & Foster (2026), Network Traffic Hashing
- Reynolds & Jacobi (2026), Memory Wipes - Performance Analysis
- Rinberg et al. (2025), Verifying LLM Inference to Detect Model Weight Exfiltration
- Scher & Thiergart (2024), Mechanisms to Verify International Agreements About AI Development (MIRI)
- Scher (2026), Verifying Restrictions on Frontier AI Research (MIRI)
- Shavit (2023), What Does It Take to Catch a Chinchilla? Verifying Rules on Large-Scale Neural Network Training via Compute Monitoring
- Teague, Rajmohan & Muehlhaeuser (2026), Could Underwater Data Centers Pose a Risk to AI Treaty Verification?, arXiv 2609.18824
- Wasil, Reed, Miller & Barnett (2024), Verification Methods for International AI Agreements
Data
- problems.csv — one row per problem
- problem_status.csv — status, scope and evidence per problem
- problem_sources.csv — the posed-by quote and other documents per problem
- leaves.csv — all sub-questions, including those without a checkable form
- leaf_sources.csv — every source row behind each sub-question