Miles Brundage is turning a long-running concern in AI governance into a new institution. After leading policy research at OpenAI for seven years, he has founded the AI Verification and Evaluation Research Institute, or AVERI, a nonprofit focused on independent safety audits of frontier AI models.
The core argument is simple: leading AI labs should not be the only parties judging whether their own systems are safe and secure. AVERI is entering that debate with money raised, a staffing target, and a proposed framework for stronger external review.
What AVERI wants to change
AVERI is built around the idea that advanced AI systems need independent evaluation, not only internal testing. Brundage left OpenAI in October 2024, where he served as an advisor on how the company should prepare for the advent of artificial general intelligence.
His concern is that major AI companies are still largely shaping their own norms. Brundage told Fortune, "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," adding, "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."
The source article notes that leading AI labs already conduct safety testing and publish technical reports. Some also work with external red team organizations. But the practical burden still falls on consumers and governments to trust what those labs choose to disclose.
That is the gap AVERI is trying to address. Independent AI safety audits would create a more formal role for outside experts, especially for frontier AI models where the consequences of weak evaluation could matter to governments, large companies and the public.
Funding shows pressure from inside the industry
AVERI has raised $7.5 million so far and is aiming for $13 million to cover 14 staff members. Its funders include former Y Combinator president Geoff Ralston and the AI Underwriting Company.
The institute has also received donations from employees at leading AI companies. Brundage described that support in blunt terms: "These are people who know where the bodies are buried," he said, "and who would like to see more accountability."
Those donations matter because they suggest that interest in independent evaluation is not coming only from outside critics. According to the source, some people inside leading AI companies are also backing a stronger accountability structure.
That does not by itself create a mandate for audits. But it does show a path for pressure to build from several directions at once: employees, funders, customers, insurers and investors.
A framework for different levels of assurance
Alongside the launch, Brundage and more than 30 AI safety researchers and governance experts published a research paper outlining a detailed framework for independent audits. The paper proposes "AI Assurance Levels" as a way to describe how strong an evaluation process is.
Level 1 roughly matches the current state described in the source article: limited third-party testing and restricted model access. At the other end, Level 4 would provide "treaty-grade" assurance robust enough to serve as a foundation for international agreements between nations.
The levels are important because the phrase independent audit can mean different things. A narrow review with little model access is not the same as an audit designed to support international confidence. By giving those differences names, the framework can help clarify what kind of evidence a company, government or buyer is actually receiving.
That distinction is especially relevant for frontier AI models. If a model is being used in important settings, the question is not only whether it has been tested, but who tested it, what access they had, and how much trust others can place in the results.
Market pressure could move faster than mandates
Brundage believes independent audits may advance even without government mandates. The source article points to several market mechanisms that could push AI companies in that direction.
Large enterprises deploying AI models for critical business processes might require audits before buying or using them. For those buyers, an independent evaluation could be a way to reduce exposure to hidden risks.
Insurance companies could also play a particularly important role. Business continuity insurers might make independent evaluations a prerequisite before writing policies for companies that rely heavily on AI. Insurers working directly with AI companies like OpenAI, Anthropic, or Google could also demand audits.
Brundage summed up that channel with a short line: "Insurance moves fast."
Investors are another potential source of pressure. If investors see independent AI audits as useful for understanding risk, they could reward companies that submit to external checks and question those that do not.
Why the audit debate matters
The launch of AVERI does not settle how independent AI safety audits should work. It does, however, give the idea a dedicated nonprofit, a funding base, and a proposed structure for comparing different degrees of assurance.
The wider question is whether AI evaluation remains mostly a company-led process or becomes a more external, standardized practice. Today, according to the source article, consumers and governments still have to rely heavily on what AI labs say about their own systems.
AVERI’s push is aimed at changing that balance. If buyers, insurers and investors begin to demand independent reviews, AI labs may have practical reasons to open their models and safety claims to more outside scrutiny.
For the future of AI governance, that could be a significant shift: from trust in company reports toward a clearer system of third-party accountability.