AVERI
AI-generated text. This page was generated using artificial intelligence.
AVERI (the AI Verification and Evaluation Research Institute) is a United States-based nonprofit focused on independent assessment of frontier AI developers' safety and security practices. It describes itself as a 501(c)(3) organization and combines pilot audits, technical research, standards development and policy advocacy. Its executive director is Miles Brundage.[1]
Auditing research
In January 2026, AVERI published Frontier AI Auditing: Toward Rigorous Third-Party Assessment of Safety and Security Practices at Leading AI Companies, a multi-author report proposing independent scrutiny of developers using secure access to non-public information. The report argues that auditing should examine organizational governance and information security as well as models, including risks arising from internal deployment.[2]
The authors propose four AI Assurance Levels to distinguish confidence in audit conclusions. They recommend AAL-1 as a baseline and AAL-2 as a near-term objective for the most advanced developers, while describing AAL-3 and AAL-4 as not yet technically and organizationally feasible. They also advocate continuing monitoring, safeguards against auditor conflicts and reports explaining scope, assumptions and limitations. These are research recommendations. The report expressly states that coauthorship does not imply endorsement of every claim or endorsement by the authors' institutions.[2]
Evaluation standards
AVERI reports that it coauthored AEF-1 with other AI Evaluator Forum members.[1] The Forum describes AEF-1, Minimum Operating Conditions for Independent Third Party AI Evaluations, as a voluntary standard addressing evaluator independence, access and transparency. Evaluators can use it to explain the conditions under which a particular assessment occurred. Its focus is the operating conditions of evaluation; it is not a statutory audit requirement or a legal compliance certificate.[3]
Policy positions
In an April 20, 2026 analysis, AVERI advocated starting mandatory audits with verification of compliance with developers' own policies, then progressing toward common standards. It favored public, redacted findings, disclosure of auditor credentials and conflicts, and flexibility if qualified-auditor capacity or standards proved inadequate. The article announced endorsement of Illinois HB 4705/SB 3261. A later editorial note identifies that proposal as a predecessor of the Illinois Artificial Intelligence Safety Measures Act (SB 315). The original legislative survey is expressly dated and should not be read as a current-status inventory.[4]
Confidential evaluation pilot
On August 27, 2026, AVERI reported a July–August pilot with Google DeepMind, OpenMined and MLCommons. It evaluated Gemini 2.5 Flash-Lite using unpublished AILuminate prompts inside a secure enclave. According to AVERI, the arrangement concealed model weights from outside evaluators and protected the benchmark prompts from collection by the developer. AVERI supplied Google DeepMind with confidential findings.[5]
The account acknowledges that the design did not address every possible means of tampering and that higher-stakes settings would need additional assurances. It also distinguishes model evaluation from a complete audit of a company's systems and operations.[5]
Funding and conflicts
As reviewed September 12, 2026, AVERI says it refuses cash donations or compensation from frontier AI companies and their executives, while disclosing API credits received from several developers. It lists philanthropic and individual funders and says no donor provides a majority of its funding. Its policies require conflict disclosures and allow recusal or declining engagements. Brundage is recused from directly auditing OpenAI until two years after his October 2024 departure, although the policy permits documented facilitation of audits led by others. These are the organization's disclosed safeguards.[1]
Related articles
- Organizations in AI Law and Policy
- United States policy on catastrophic AI risk
- Illinois Artificial Intelligence Safety Measures Act (SB 315)
- Google DeepMind
External links
References
- ↑ 1.0 1.1 1.2 AVERI, About, accessed September 12, 2026.
- ↑ 2.0 2.1 Miles Brundage and coauthors, Frontier AI Auditing: Toward Rigorous Third-Party Assessment of Safety and Security Practices at Leading AI Companies, January 15, 2026.
- ↑ AI Evaluator Forum, Minimum Operating Conditions for Independent Third Party AI Evaluations, accessed September 12, 2026.
- ↑ Miles Brundage, Frontier AI Auditing-Related Legislation in the US: Landscape, Challenges, and a Path Forward, April 20, 2026, with subsequent editorial note; accessed September 12, 2026.
- ↑ 5.0 5.1 Carly Tryens, AVERI Pilot Report: The World’s First Double-Blind Evaluation of a Proprietary Language Model, August 27, 2026.