P021 Trustworthiness and leakage of monitoring sensors and data
Can facility sensors be built that both parties trust — verifier trusts the readings, prover trusts they do not exfiltrate — with a quantified leak bound?
Whether facility sensors and their data can be trusted and leak little.
Status
Started 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
- Harack et al. (2025), Verification for International AI Governance (Oxford AIGI):
these data streams can bring their own potential vulnerabilities, so they need to be considered carefully
(§2.2.4.1, p.48, L2634–2635) (source states the problem but not the quantity the question asks for)
Also raised in: Aarne & Petrie (2025), International Security Applications of Flexible Hardware-Enabled Guarantees (flexHEG Part 3); Abdelghafar & Kulp (2026), Privacy-Preserving AI Verification via Minimal Information Disclosure; Proofworks, Living Verification Documents.
Sub-questions
| ID | Sub-question | What the sources report |
|---|---|---|
| P021.1 | Has a facility sensor suite been threat-modelled with a quantified leak bound (bits/s) per channel, showing sensors cannot exfiltrate or sabotage? | open |
| P021.2 | Can workload type be inferred from rack video (LEDs, fan speed) above chance? | open |
| P021.3 | Given a sensor's time resolution and a verifier-triggered external pulse, what location error (km) results, and what time-coarsening defeats triangulation? | open |
| P021.4 | Does aggregated or allocation-level facility power data leak computation details beyond what is declared? | open |
| P021.5 | Has a privacy-preserving protocol (MPC, confidential computing, statistical) been demonstrated that checks a declared grid load list against verifier estimates and bounds undeclared MW without revealing load profiles? | open |
Sub-questions not yet stated precisely enough to check
| ID | Sub-question |
|---|---|
| P021.6 | Can analog sensors be built that both parties trust (verifier trusts readings; prover trusts no leakage)? |
| P021.7 | How are monitoring systems that must themselves be air-gapped in turn monitored? |
| P021.8 | Should AI infrastructure be kept away from sensitive military assets to limit what monitoring reveals? |
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 P021.1: Has a facility sensor suite been threat-modelled with a quantified leak bound (bits/s) per channel, showing sensors cannot exfiltrate or sabotage?
For P021.2: Can workload type be inferred from rack video (LEDs, fan speed) above chance?
For P021.3: Given a sensor's time resolution and a verifier-triggered external pulse, what location error (km) results, and what time-coarsening defeats triangulation?
For P021.4: Does aggregated or allocation-level facility power data leak computation details beyond what is declared?
For P021.5: Has a privacy-preserving protocol (MPC, confidential computing, statistical) been demonstrated that checks a declared grid load list against verifier estimates and bounds undeclared MW without revealing load profiles?
For P021.6: Can analog sensors be built that both parties trust (verifier trusts readings; prover trusts no leakage)?
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