P069 Hardware enforcement of permitted use
Can chips be restricted to approved code, inference-only, or bounded training, and what throughput penalty does the best circumvention attack achieve?
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.
Status
Analyzed 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
- Reuel, Bucknall et al. (2025), Open Problems in Technical AI Governance, compute sections:
How could the use of high-end chips in training foundation models be prevented?
(§6.2, p.34, L1883 (Q79)) (source states the problem but not the quantity the question asks for)
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 (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); Hausenloy & Li (2026), The Compute Verification Post; Kulp et al. (2024), Hardware-Enabled Governance Mechanisms (RAND WR-A3056-1); Wasil, Reed, Miller & Barnett (2024), Verification Methods for International AI Agreements; AI Security Forum, Requests for Discussion; Amodo (2026), AI 2040 Plan A — How You Can Help; Dean, Verification Plan (AI 2040); Proofworks, Living Verification Documents.
Sub-questions
| ID | Sub-question | What the sources report |
|---|---|---|
| P069.1 | What training-throughput penalty does the best known attack achieve when training on inference-specialised chips, or on chips converted to inference-only via firmware/licences? | open |
| P069.2 | Has any prototype refused to execute a training job deviating from a signed plan, in a red-team test? | open |
| P069.3 | Can a firmware cluster-size cap (e.g. no training across more than 16 H100-equivalents) be bypassed by low-communication training or modified drivers while staying under the cap? | open |
| P069.4 | What false-positive and false-negative rates does a chip-side 'unauthorised training' detector / flexHEG workload classifier achieve on a benchmark workload set against disguise? | open |
| P069.5 | What throughput overhead does a firmware reporting mechanism impose on a representative training job? | open |
| P069.6 | Can a single FIB edit remove a clock-throttle on a test chip, at what cost? | open |
| P069.7 | How many weeks does it take to convert a stated share of global compute to verified inference-only? | open |
| P069.8 | Can a 10^24 FLOP training run proceed on a cluster with scale-out interconnect removed (crude first-year modification)? | open |
| P069.9 | Can hardware or software restrict chips to approved code/models or per-workload sign-off without algorithm access, given that software alone can verify but not enforce? | a source suggests an answer (not checked) |
| P069.10 | Can a hardware license mechanism distinguish licensed from unlicensed inference, and give the licensor feedback on use (offline licensing is open-loop)? | open |
Sub-questions not yet stated precisely enough to check
| ID | Sub-question |
|---|---|
| P069.11 | Can chips with hardcoded weights serve as on-device inference-only verification hardware? |
| P069.12 | Can an operation-stack disallow-list (e.g. blocking backpropagation) be enforced? |
| P069.13 | Which functionality should be throttled so dual-use workloads are hit while non-AI uses are spared, and can that line be drawn? |
| P069.14 | Has a commercial data centre of stated size run inference-only verification for a stated duration with third-party checking? |
| P069.15 | Can caps on technical inputs, vetoes on frontier-pushing runs, or enduring rulesets be implemented technically? |
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 P069.5: What throughput overhead does a firmware reporting mechanism impose on a representative training job?
For P069.6: Can a single FIB edit remove a clock-throttle on a test chip, at what cost?
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