P060 Weight snapshots on accelerators
Can an accelerator take attested, tamper-evident weight/memory snapshots bound to chip state at trigger time, at what overhead and with what storage cost?
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.
Status
Analyzed 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
- Harack et al. (2025), Verification for International AI Governance (Oxford AIGI):
it depends on speculative weight snapshotting hardware features that are not yet available
(§2.1.3 fn 131, p.41, L2268) (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); 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); AI Security Forum, Requests for Discussion; Proofworks, Living Verification Documents.
Sub-questions
| ID | Sub-question | What the sources report |
|---|---|---|
| P060.1 | Does any shipping accelerator support attested, signed, tamper-evident on-chip weight/memory snapshots at a specified interval and overhead? | open |
| P060.2 | Can host software predict a random or verifier-triggered snapshot early enough to swap HBM contents, and can a protocol bind the snapshot to chip state at trigger time? | partly answered |
| P060.3 | Do current accelerators expose non-volatile storage writable only by signed firmware and physically readable by a verifier (workload logs)? | implied by a source, not stated as open |
| P060.4 | Can a GPU answer a timed challenge (<=50 ms) only with the declared weights resident, at a stated serving overhead (a target to be agreed)? | open |
| P060.5 | Can a current accelerator remotely attest the hash of the weights loaded for a serving job, at what overhead? | open |
| P060.6 | Can a system-software protocol locate model weights and data on a cluster when a red team obfuscates them (e.g. weights hidden in benign outputs), at a stated detection rate? | open |
| P060.7 | What is the time to stream and hash-check a 1 TB model through a secure-boot gateway at a stated bandwidth, and what attack surface does it add? | open |
| P060.8 | What throughput loss does random HBM snapshot+hash cause on a large-scale (stated accelerator count) synchronous job? | open |
| P060.9 | Can firmware recover weight regions from logged compiled code plus allocation records for standard sharded stacks? | open |
| P060.10 | What storage (and cost) does retaining full transcripts / gradients / checkpoints require for a frontier run, and what is lost when checkpoints are deleted? | partly answered |
Sub-questions not yet stated precisely enough to check
| ID | Sub-question |
|---|---|
| P060.11 | Can monitoring weight changes in memory reliably detect training? |
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 P060.1: Does any shipping accelerator support attested, signed, tamper-evident on-chip weight/memory snapshots at a specified interval and overhead?
For P060.4: Can a GPU answer a timed challenge (<=50 ms) only with the declared weights resident, at a stated serving overhead (a target to be agreed)?
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