Former OpenAI Executive Launches Independent Institute to Audit AI Safety Claims
San Francisco, Saturday, 24 January 2026.
Miles Brundage has secured $7.5 million to establish AVERI, challenging AI companies’ self-regulation practices and pushing for mandatory third-party safety audits.
AVERI’s Mission and Funding Structure
The AI Verification and Evaluation Research Institute (AVERI), announced on January 23, 2026, positions itself as a think tank rather than a direct auditing firm [1]. The nonprofit has raised $7.5 million toward a $13 million funding goal to support 14 staff members and cover two years of operations [1]. The funding comes from diverse sources including Halcyon Futures, Fathom, Coefficient Giving, Geoff Ralston, Craig Falls, Good Forever Foundation, Sympatico Ventures, and the AI Underwriting Company [1]. AVERI’s motto, “Making third-party auditing of frontier AI effective and universal,” reflects its ambitious scope in addressing what Brundage identifies as a critical gap in AI oversight [2].
Addressing Industry Self-Regulation Concerns
Brundage’s initiative stems from his seven-year tenure at OpenAI, where he observed companies developing their own safety standards without external oversight [1]. “One of the things I learned while working at OpenAI is that companies are figuring out the norms of this kind of thing on their own. There’s no one forcing them to work with third-party experts to make sure that things are safe and secure. They kind of write their own rules,” Brundage explained [1]. After leaving OpenAI in October 2024, Brundage concluded that the AI industry requires independent verification similar to other high-stakes sectors [1]. The institute’s research framework, published alongside AVERI’s launch and co-authored by over 30 AI safety researchers and governance experts, addresses the fundamental challenge that most current third-party AI assessments rely solely on public information, which proves insufficient for meaningful safety assurance [3].
Structured Audit Framework and Implementation Timeline
AVERI’s technical approach centers on AI Assurance Levels (AAL), ranging from AAL-1 to AAL-4, with increasing levels of rigor and cost [3]. AAL-1, representing limited assurance, involves time-bounded assessments using API access and some non-public information, with estimated costs between $300,000 and $600,000 per engagement [3]. AAL-2, targeting moderate assurance over several months with gray-box system access and staff interviews, could cost $1 million or more [3]. The higher levels, AAL-3 and AAL-4, involve continuous oversight and comprehensive access, potentially costing several million dollars annually [3]. The framework anticipates AAL-2 audits becoming feasible in early to mid-2026, while AAL-3 could be possible by early 2027 if significant investments begin promptly [3]. The most rigorous AAL-4 audits, designed to detect active deception attempts, might be necessary by late 2027 for verifying international AI safety cooperation [3].
Market Adoption and Regulatory Alignment
AVERI’s strategy relies on multiple stakeholders to drive adoption of independent AI audits. Insurance companies and investors represent key potential drivers, as they require standardized risk data to make informed decisions about AI-related coverage and investments [1][3]. The European Union’s AI Act, which recently entered force, provides regulatory momentum by suggesting external evaluations for models posing “systemic risks” through its Code of Practice for General Purpose AI [1]. However, the Act stops short of explicitly mandating audits of companies’ evaluation procedures [1]. Brundage envisions building “dream teams” combining expertise from traditional audit firms, cybersecurity companies, AI safety nonprofits, and academic institutions to address the current shortage of qualified AI auditors [1]. The institute’s goal extends beyond technical evaluation to encompass what Brundage describes as achieving “a level of scrutiny that is proportional to the actual impacts and risks of the technology, as smoothly as possible, as quickly as possible, without overstepping” [1].