Miles Brundage’s OpenAI Exit Led to an AI Audit Campaign

Miles Brundage left his position as OpenAI’s Senior Advisor for AGI Readiness in October 2024, warning of substantial gaps in the preparedness of frontier laboratories and society. His October 23 departure statement set out his reasons for moving outside the industry and disclosed that OpenAI would distribute the remaining AGI Readiness staff and projects among other teams.

The resignation is a historical event, not a new departure. The consequential update came in January 2026, when Brundage formally launched the AI Verification and Evaluation Research Institute, or AVERI; its launch announcement identifies the organization as a US-based nonprofit pursuing rigorous third-party scrutiny of frontier developers’ safety and security claims.

The warning was about readiness, not an AGI breakthrough

Brundage’s assessment did not mean that OpenAI had secretly crossed a universally recognized threshold for artificial general intelligence. It concerned whether companies and public institutions could safely develop, deploy and govern systems as their capabilities increased.

That distinction matters because AGI has no single, generally accepted operational test. A laboratory can face consequential questions about model evaluations, cybersecurity, deployment controls, economic disruption and regulatory oversight before its systems satisfy any particular definition of human-level intelligence.

The warning also distinguished current conditions from future progress. Declaring institutions unprepared at a particular moment does not prove that they can never become prepared; it identifies gaps that must be closed before more capable systems make those weaknesses harder or more costly to address.

Why Brundage moved outside the company

The decision reflected several connected constraints rather than one public dispute. Work inside a prominent AI developer limited the range and speed of research Brundage could publish, while the company’s commercial interests complicated the perceived independence of policy proposals originating from within it.

External work also offered a broader field of action. Questions such as frontier-model regulation, compute governance, economic effects and cross-company safety standards cannot be resolved by one developer, even when that developer improves its own controls.

The departure was not framed as a rejection of every safety effort at OpenAI. It represented a judgment about where industry-wide policy research and advocacy could operate with greater independence, especially when the subject includes the incentives and conduct of OpenAI itself.

The AGI Readiness team ceased to be a standalone unit

The immediate organizational result was the redistribution of the team’s work. Economic research moved into OpenAI’s economics function, while other personnel and projects transferred elsewhere within the company.

This is narrower than claiming that all preparedness work stopped. The dedicated unit disappeared, but at least some responsibilities continued under different management. Public information alone does not establish whether that structure ultimately improved integration or weakened accountability.

A centralized team gives preparedness work a visible mandate and a recognizable route for escalation. A distributed model can place risk analysis closer to product, research and security decisions, but it may also make budgets, authority and ownership more difficult for outsiders to follow.

AVERI converts the warning into an audit proposal

AVERI is the clearest concrete development after the resignation. Its objective is to help make third-party auditing of frontier AI effective and broadly adopted, with reviewers assessing developers’ systems and internal practices against relevant standards.

This model goes beyond testing a public chatbot. Meaningful scrutiny may require controlled access to non-public evaluations, safeguards, development processes and organizational records. Auditors therefore need enough evidence to challenge corporate claims without exposing sensitive model details, security measures or commercially valuable information.

AVERI is positioned primarily as a research institute working on the design, incentives and infrastructure for audits, rather than as a conventional certification business issuing a simple pass-or-fail badge. That leaves important questions unresolved, including who sets the standards, how reviewer competence is measured and how financial conflicts are managed.

The business relevance extends beyond debates over AGI terminology. Enterprise customers, insurers, investors and public-sector buyers may all need evidence that a supplier’s controls function as represented. Comparable independent assessments could make frontier-AI risk easier to evaluate, provided auditors receive meaningful access and use standards capable of revealing weaknesses.

OpenAI has expanded its internal governance documentation

In May 2026, OpenAI published a Frontier Governance Framework covering risk assessment, model reporting, incident response, external expert input and threats involving cyber offense, chemical or biological misuse, harmful manipulation and loss of control; the document also places the company’s Preparedness Framework at the foundation of its approach to severe risks.

The publication demonstrates that preparedness work did not vanish with the former team. It does not independently prove that the underlying controls are adequate or consistently followed. A company-authored framework documents processes and commitments, while an external audit would test those commitments against evidence and an independently defined standard.

Those mechanisms are not inherently opposed. Internal governance is necessary for day-to-day decisions, and outside scrutiny can test whether that governance operates as intended. The unresolved issue is the quality of the evidence available to reviewers and whether developers permit enough access for scrutiny to be meaningful.

The readiness question remains open

No generally accepted test demonstrates that OpenAI, its competitors or society has reached a final state of AGI readiness. Technical safeguards, corporate authority, emergency procedures, economic policy and public regulation develop at different speeds and cannot be reduced to one corporate score.

Brundage’s move sharpened the institutional conflict at the center of frontier-AI governance. Developers possess the most detailed knowledge of their systems, but their commercial incentives create a reason for independent verification. Outside reviewers offer distance from those incentives, yet their conclusions are only as strong as their access, standards and expertise.

The story has therefore moved beyond one executive adviser’s departure. OpenAI has documented a broader internal governance structure, while its former readiness adviser has built an institution intended to make external scrutiny practical. Whether those safeguards work remains a question for verifiable evidence, not the existence of a framework alone.

Miles Brundage Left OpenAI; His Readiness Warning Became an Audit Campaign