Private technical workshop · Berkeley, California
A workshop on the technical tooling needed to verify claims about frontier AI development — what should be proven, to whom, and how.
Attendance by invitationWhy now
“Today, the world lacks the technical and governance tools to deliberately pace frontier-wide progress.”
Pacing the Frontier — a statement signed by 1,386 employees of frontier AI companies, July 2026
There is increasing agreement that independent assurance is valuable for verifying that AI developers' safety commitments are upheld and that risks from AI development are managed effectively. While embedding trusted auditors in AI companies represents a step toward such assurance, it is currently unclear how such a model could adapt to support international coordination on pacing AI progress. Technical tooling that enables trustworthy, automated, and privacy-preserving verification could help to address this gap.
This workshop aims to bring together auditors, external researchers building verification mechanisms, and staff from AI developers to identify and prioritize what tooling needs to be developed.
Focus
The workshop centers on concrete claims that technical mechanisms could help establish — for example, which models are deployed, whether specified safeguards remain in place, how AI compute is being used, and whether agreed constraints on development or deployment are being followed. Participants will examine the feasibility and limits of existing approaches across relevant use-cases including internal governance and security, domestic auditing, and international coordination. They will stress-test proposals on near-term priorities and work toward a shared view of what to build next.
A short, shared list of claims about AI safety and pacing for which stronger verification would be most valuable, in domestic and international contexts.
Research questions and pilot projects that AI developers, external researchers, and other stakeholders could advance in the next 6–18 months.
Program at a glance
Sessions and timings may still shift; participants will receive final details ahead of the event.
Workshop goals and program overview.
Which claims we want to verify, why, and the technical approaches being developed, with Q&A.
Short presentations on selected verification approaches, followed by discussion.
Cryptographic approaches · retrofitted devices · China and middle powers · verification integrity and tampering.
Groups share the most promising experiments and proposed roles for external researchers and AI developers.
Prioritizing ideas for collaboration · making verification research easier for labs to engage with · shared infrastructure and public goods · what US policy-makers would need to rely on verification · where AI could automate verification R&D.
Pre-reading
Proposes a retrofittable reference architecture separating evidence capture from later evaluation of randomly challenged workload records.
Proposes a phased verification plan for an international AI slowdown agreement, combining compute accounting, mutual declarations, inference-only retrofits, network taps, and partial recomputation.
Proposes hardware-independent verification using prover isolation, cryptographic commitments, and challenge-based recomputation to bound unexplained information in network outputs.
Describes the performance of Attestable's system for zero-knowledge proofs of inference on open-source models with 30–70B parameters.
For a more comprehensive reading list, see this document.
Who's in the room
Around thirty participants, spanning the communities whose combined judgment the problem requires:
Research, infrastructure, security, and policy staff from leading AI companies working on or adjacent to verification.
Organizations building mechanisms for verifying claims about AI compute and workloads.
Third parties developing auditing and risk evaluation practices that verification infrastructure could support.
Philanthropic and institutional supporters positioned to back the projects the workshop prioritizes.
Format