The AI Verification and Evaluation Research Institute (AVERI) is a US-based 501(c)(3) non-profit research institute whose stated mission is "to make frontier AI auditing effective and universal" — third-party verification of leading AI developers' safety and security claims, and evaluation of their systems and practices against relevant standards (Source: averi.org). Its executive director is Miles Brundage, formerly of OpenAI; Sean McGregor is a co-founder. The organization launched publicly in 2026 alongside a report on frontier AI auditing co-authored with contributors at dozens of organizations (Source: milesbrundage.substack.com).
AVERI's framing is that safety and security claims about AI systems are largely self-reported by companies, and that independent verification requires "deep, secure access to non-public information — access that developers aren't currently required to provide." It draws the analogy to aviation, finance, and nuclear power as industries where external oversight displaced self-reporting (Source: averi.org).
Activities
AVERI describes itself as "a standard-setter grounded in practice," running four lines of work.
Pilot audits. Frontier AI audits are not currently required, and AVERI conducts voluntary pilots with companies that agree to them, choosing areas other organizations are not already pursuing. It stated in 2026 that further pilot results would follow in the third quarter of that year (Source: averi.org).
Auditing standards. AVERI co-authored AEF-1, a standard defining minimum operating conditions for independent third-party evaluations, with other members of the AI Evaluator Forum (Source: averi.org).
Policy analysis and advocacy. AVERI provided feedback on and endorsed the legislation that became Illinois SB 315 (frontier safety framework with mandatory third-party audits), which it describes as the first US frontier AI audit requirement, taking effect in 2028 (Source: averi.org).
Open-source tooling. Among its first projects, AVERI states it is developing privacy-preserving audit automation tools that let auditors analyze sensitive non-public information while safeguarding confidentiality (Source: averi.org).
Double-blind evaluation pilot
AVERI was a partner in the double-blind evaluation pilot published on August 27, 2026 with Google DeepMind, MLCommons, OpenMined and the Singapore AI Safety Institute, in which Gemini 2.5 Flash Lite was evaluated against reserve-set MLCommons AILuminate prompts inside a hardware-encrypted enclave. Four AVERI staff are listed among the technical report's authors — Sean McGregor, George Balston, Patricia Paskov and Miles Brundage. AVERI encrypted and decrypted the prompts and outputs, and AVERI staff evaluated the outputs. The report records that not all of the model owner's inference code could be inspected or allow-listed, and that "AVERI was informed of this challenge and accepted the overall setup with the enclave" (Double Blind Evals: Resolving the Dual Confidentiality Dilemma in AI Safety Auditing). AVERI published its own account of the pilot as its first pilot findings (Source: averi.org).
Funding and independence
AVERI states that no donor makes up a majority of its funding, and names as funders the AI Underwriting Company, Coefficient Giving, Craig Falls, Geoff Ralston, the Good Forever Foundation, Fathom, Halcyon Futures, Sympatico Ventures, "and several non-executive employees and alumni of frontier AI companies." It has received API credits as donations or compensation from Amazon, Anthropic, Google DeepMind, Microsoft, OpenAI and Thinking Machines Lab, and states that it does not accept cash donations or compensation from frontier AI companies or their executives (Source: averi.org).
The organization publishes a conflict-of-interest policy requiring employees, board members and contractors to disclose potential conflicts on starting and biannually, with mitigation by recusal, declined engagement, external peer review, heightened oversight, or transparency statements in audit reports. Its most significant disclosure concerns Brundage, who earned equity in OpenAI while employed there, has sold all shares he is eligible to sell, and is recused from directly auditing OpenAI until at least two years after his departure — though he may help facilitate an audit principally led by others, with his involvement documented in related public outputs (Source: averi.org).
That developer-pays and conflict question is the recurring criticism of the third-party certification model surveyed at Independent Verification Organizations (IVOs); AVERI's stated answer is to refuse cash from the audited class entirely and to publish the residual exposures.
Relationships
- instance-of: Independent Verification Organizations (IVOs) — a third-party verifier of developer safety claims.
- related: Double-Blind Evaluation — partner in the first pilot of the framework.
- related: Double Blind Evals: Resolving the Dual Confidentiality Dilemma in AI Safety Auditing — technical report co-authored by AVERI staff.
- related: Miles Brundage — executive director.
- related: Illinois SB 315 (frontier safety framework with mandatory third-party audits) — legislation AVERI endorsed and gave feedback on.
- related: METR — another organization running voluntary frontier evaluations.
- related: AI Pre-Release Vetting, Alignment Auditing.