P055 Keys and hardware roots of trust
What lab cost and time are needed to extract an accelerator's device/attestation key, and can keys be provisioned so neither party holds a copy?
Extraction resistance and provisioning of device keys and roots of trust.
Terms: root of trust (RoT): a small, highly trusted hardware component that the rest of a system's security is built on. HSM (hardware security module): a tamper-resistant device that stores and uses cryptographic keys. Caliptra, OpenTitan: open hardware root-of-trust designs.
Status
Posed
Posed by
- Harack et al. (2025), Verification for International AI Governance (Oxford AIGI):
there are serious ongoing questions about the security of on-chip keys of various kinds
(§2.2.2.1, p.44, L2396) (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); Petrie, Aarne, Ammann & Dalrymple (2025), Overview of Flexible Hardware-Enabled Guarantees (flexHEG Part 1); 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); Shavit (2023), What Does It Take to Catch a Chinchilla? Verifying Rules on Large-Scale Neural Network Training via Compute Monitoring; Kulp et al. (2024), Hardware-Enabled Governance Mechanisms (RAND WR-A3056-1); AI Security Forum, Requests for Discussion; Proofworks, Living Verification Documents.
Sub-questions
| ID | Sub-question | What the sources report |
|---|---|---|
| P055.1 | What cost and time does a well-equipped lab need to extract the device identity / attestation key from an H100/B200-class accelerator by physical attack (or defeat signed-firmware enforcement)? | partly answered |
| P055.2 | How many traces / hours does the best side-channel (power/EM) attack need to extract the attestation key from an accelerator RoT or a HEM guarantee processor? | implied by a source, not stated as open |
| P055.3 | Once an attack on a mechanism is known, what is its marginal per-chip cost, time and success rate — i.e. does it scale — and can moderate per-chip fault-injection detection suffice statistically? | implied by a source, not stated as open |
| P055.4 | Has any current accelerator's key storage or security module undergone an external physical-attack evaluation at a high Common Criteria / FIPS 140-3 level, and is HSM/SEP evidence transferable to state attackers? | implied by a source, not stated as open |
| P055.5 | Is there a device-key provisioning process witnessed by both parties such that neither can hold a copy of the key? | open |
| P055.6 | Can firmware-signing authority on a current accelerator be transferred from the vendor to a Verifier by key revocation? | partly answered |
| P055.7 | Does any leading AI accelerator include a dedicated hardware security module region? | implied by a source, not stated as open |
| P055.8 | Does any shipping accelerator with significant data-centre share integrate an open-source, externally audited RoT (Caliptra/OpenTitan) whose silicon a third party checked against the design? | partly answered |
| P055.9 | Has an open-source design providing secure boot + confidential computing + on-chip tamper response (or a guarantee-processor IP block) been taped out and passed external red-team / glitch / side-channel testing? | open |
Sub-questions not yet stated precisely enough to check
| ID | Sub-question |
|---|---|
| P055.10 | Could one software/firmware vulnerability, or a covert compromise plus withheld signatures, give fast scalable compromise across a fleet? |
| P055.11 | Can security rest on something other than stored hardware keys (e.g. secret bits embedded in physical matter)? |
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 P055.1: What cost and time does a well-equipped lab need to extract the device identity / attestation key from an H100/B200-class accelerator by physical attack (or defeat signed-firmware enforcement)?
For P055.2: How many traces / hours does the best side-channel (power/EM) attack need to extract the attestation key from an accelerator RoT or a HEM guarantee processor?
For P055.3: Once an attack on a mechanism is known, what is its marginal per-chip cost, time and success rate — i.e. does it scale — and can moderate per-chip fault-injection detection suffice statistically?
For P055.6: Can firmware-signing authority on a current accelerator be transferred from the vendor to a Verifier by key revocation?
For P055.8: Does any shipping accelerator with significant data-centre share integrate an open-source, externally audited RoT (Caliptra/OpenTitan) whose silicon a third party checked against the design?
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