P096 Inference deployment monitoring

Can an operator prove all inference passed through a given monitor, and how well can misuse or diversion be caught among approved traffic?

Monitoring inference deployments for coverage, misuse and diversion.

Status

Analyzed 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.
Scher & Thiergart (2024), Mechanisms to Verify International Agreements About AI Development (MIRI; arXiv June 2025) (2024-11), L5548–5550

Posed by

Also 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, Ammann & Dalrymple (2025), Overview of Flexible Hardware-Enabled Guarantees (flexHEG Part 1); Aarne & Petrie (2025), International Security Applications of Flexible Hardware-Enabled Guarantees (flexHEG Part 3); Cankaya (2026), A System Overview for Near-Term, Low-Trust AI Compute Verification (v0.2); 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); Hausenloy & Li (2026), The Compute Verification Post; Proofworks, Living Verification Documents.

Sub-questions

IDSub-questionWhat the sources report
P096.1Can an operator prove to an outside inspector that all inference requests on a cluster passed through a given classifier/monitor, with a measured coverage gap?open
P096.2Can attested inference receipts show compute spent on monitors/safeguards vs the policy model to a stated precision?open
P096.3At a stated false-positive rate, what recall does a compliance classifier achieve on blacklisted-use tasks hidden among whitelisted-model traffic?open
P096.4What compute overhead and latency do separate verification infrastructures and compliance test suites add to serving?open
P096.5Can inference weight containment (all instances known) be made internationally verifiable and insider-resistant, e.g. via bottlenecked distributed inference?open
P096.6What is the attack success rate of prompt injection against LLM inspector agents, and does classifier robustness keep pace with AI progress?open
P096.7Can agent computations using external tools (web) be logged and replayed for verification?open
P096.8What is the latency from the start of an above-threshold run to an alarm, for a given mechanism?open
P096.9What fraction of a cluster's GPU-hours can be diverted before scheduler/DCGM telemetry flags it?implied by a source, not stated as open

Sub-questions not yet stated precisely enough to check

IDSub-question
P096.10What share of deployment usage must be examined to catch small, causally unrelated defections, and how can the examined space be limited?
P096.11Does deployment governance require more sensitive information than development governance?
P096.12Is detailed verification for computational emergencies workable?

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 P096.6: What is the attack success rate of prompt injection against LLM inspector agents, and does classifier robustness keep pace with AI progress?

Data