P023 Residual covert-channel capacity out of monitored compute
How many bits per second or per token survive monitoring via steganography, protocol free variables, and physical channels (fibre, coolant, RF)?
Measured bits/s or bits/token that survive monitoring: steganography in outputs and egress, protocol free variables, and physical channels (fibre, coolant, RF).
Status
Started 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
- Harack et al. (2025), Verification for International AI Governance (Oxford AIGI):
Even if you can inspect all data passing into or out of a data center, you won’t know its full meaning.
(§2.5.2.4, p.64, L3482–3483) (source states the problem but not the quantity the question asks for)
Also 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); Cankaya (2026), A System Overview for Near-Term, Low-Trust AI Compute Verification (v0.2); 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), 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); Canary Institute research; AI 2040 (2026), Get Involved in Verification; AI Security Forum, Requests for Discussion; Proofworks, Living Verification Documents.
Sub-questions
| ID | Sub-question | What the sources report |
|---|---|---|
| P023.1 | What covert data rate has been demonstrated via hidden or tapped optical fibre (DAS, extra wavelengths) across a rack boundary (bps)? | open |
| P023.2 | What acoustic data rate has been demonstrated along a data-centre coolant loop (bps)? | open |
| P023.3 | What is the measured capacity of a conducted signal through a stated length of treated coolant (bps)? | open |
| P023.4 | What covert rate does RF emanation from stock (unmodified) rack hardware give at 3 m (bps)? | partly answered |
| P023.5 | What training-efficiency penalty (loss per FLOP) does low-communication training incur at a stated covert inter-pod bandwidth (kbps)? | open |
| P023.6 | What residual covert or unexplained egress capacity (bit/s per server or pod) survives an active-warden NS tap on real edge traffic under red team (validating the ~10^2 bps Fermi estimate, a covert-capacity estimate from internal notes, or the <=10 KB/s per pod bound)? | partly answered |
| P023.7 | How long does a covert training run take if covert output is capped at <=10 KB/s per pod, i.e. is that cap enough to heavily penalise covert training? | partly answered |
| P023.8 | How many bits per output token (or per checkpoint) can be hidden steganographically while passing exact or near-exact replay checks (e.g. Token-DiFR)? | partly answered |
| P023.9 | How many covert bits per packet remain in TCP/IP transport headers after header scrubbing? | partly answered |
| P023.10 | How many covert bits/s remain in canonicalised egress 'free variables' (diagnostic heartbeats, ns timestamps) at a fixed rate? | partly answered |
| P023.11 | What timing-channel capacity remains after padding inputs and outputs at the inline capture point? | partly answered |
| P023.12 | How many bits/day leak through the selection, order or timing of hash publication under fixed-order, fixed-interval batching? | partly answered |
| P023.13 | What fraction of hashed API responses can be inverted by dictionary lookup under a given slicing scheme? | partly answered |
| P023.14 | How many bytes per hash minimise verifier storage while keeping dictionary-attack success below a stated bound? | open |
| P023.15 | How much leaks through undisclosed wavelength modulation or timing on a split egress fibre with spectrum monitoring? | open |
| P023.16 | What bit rate can an NVLink-contention covert channel achieve between co-located workloads on one node? | open |
| P023.17 | Under physical controls (RF shielding / per-pod Faraday cages, cameras), what covert bandwidth can a red team get into or out of a pod outside the tapped links? | open |
| P023.18 | How many bits can be hidden in mid-training weights/optimizer state undetected by re-execution of later steps? | open |
Possibly relevant work found by automated search
Found by AI agents searching for this problem's sub-questions, mostly by keyword and citation in the OpenAlex scholarly database and on the web; a second AI pass reviewed each one, and we spot-checked some by hand. They seemed reasonably related, as far as one can tell from outside each paper's field. Some come from neighbouring fields, such as arms control and financial audit, and bear on the problem by analogy. Not reviewed against the status above; a paper listed here does not change that status. 928 of 1052 quotes checked word for word against the paper's text.
For P023.1: What covert data rate has been demonstrated via hidden or tapped optical fibre (DAS, extra wavelengths) across a rack boundary (bps)?
For P023.2: What acoustic data rate has been demonstrated along a data-centre coolant loop (bps)?
For P023.3: What is the measured capacity of a conducted signal through a stated length of treated coolant (bps)?
For P023.4: What covert rate does RF emanation from stock (unmodified) rack hardware give at 3 m (bps)?
For P023.5: What training-efficiency penalty (loss per FLOP) does low-communication training incur at a stated covert inter-pod bandwidth (kbps)?
For P023.6: What residual covert or unexplained egress capacity (bit/s per server or pod) survives an active-warden NS tap on real edge traffic under red team (validating the ~10^2 bps Fermi estimate, a covert-capacity estimate from internal notes, or the <=10 KB/s per pod bound)?
For P023.8: How many bits per output token (or per checkpoint) can be hidden steganographically while passing exact or near-exact replay checks (e.g. Token-DiFR)?
For P023.11: What timing-channel capacity remains after padding inputs and outputs at the inline capture point?
For P023.13: What fraction of hashed API responses can be inverted by dictionary lookup under a given slicing scheme?
For P023.16: What bit rate can an NVLink-contention covert channel achieve between co-located workloads on one node?
For P023.17: Under physical controls (RF shielding / per-pod Faraday cages, cameras), what covert bandwidth can a red team get into or out of a pod outside the tapped links?
For P023.18: How many bits can be hidden in mid-training weights/optimizer state undetected by re-execution of later steps?
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