Wiki · Person · Last reviewed May 20, 2026

Miles Brundage

Miles Brundage is an AI policy researcher and nonprofit leader known for OpenAI policy and AGI readiness work, research on verifiable claims and frontier AI governance, and the founding leadership of AVERI, a nonprofit focused on third-party auditing of advanced AI systems.

Snapshot

OpenAI Policy and AGI Readiness

Brundage joined OpenAI in 2018, before ChatGPT made frontier AI governance a mainstream political issue. His work there sat in the early policy layer around deployment, safety documentation, red teaming, preparedness, and the social consequences of increasingly capable models.

OpenAI's public archive lists Brundage on work about verifiability, agentic AI systems, frontier AI regulation, and safety approaches for advanced models. That record matters because it shows how AI policy inside a frontier lab became more than public relations. It involved concrete questions about what a developer should measure, disclose, withhold, audit, and gate before releasing a more capable system.

In October 2024, Brundage left OpenAI and wrote that his work would be more effective from outside the company. Reputable reporting at the time described him as a senior AGI readiness adviser and policy leader, and noted his concern that neither OpenAI, other frontier labs, nor the world were ready for AGI. The durable significance is not one resignation event. It is the shift from internal readiness advising to independent pressure for stronger external governance.

Verifiable Claims

Brundage is the first-listed author of Toward Trustworthy AI Development: Mechanisms for Supporting Verifiable Claims, a 2020 report arguing that AI developers need ways to make safety, security, fairness, privacy, and development-process claims that outsiders can check.

The paper is important because it moved beyond abstract principles. Instead of saying that AI systems should be trustworthy, it asks what evidence would let customers, governments, civil society, employees, and other stakeholders verify that a claim is true. It discusses institutional mechanisms, software mechanisms, hardware mechanisms, audits, documentation, privacy-preserving analysis, and controlled access to sensitive information.

That frame has become more central as frontier companies publish system cards, preparedness frameworks, responsible-scaling policies, model cards, and safety reports. A document is not the same as evidence. Brundage's verifiable-claims work asks how documentation becomes accountable rather than merely decorative.

Governance Research

Brundage's research record spans several recurring governance problems. The Malicious Use of Artificial Intelligence, published in 2018 with a large group of authors, helped make AI-enabled cyber, physical, and information harms a policy topic before generative AI became public infrastructure.

Later work on frontier AI regulation framed advanced foundation models as a distinct governance target: systems with capabilities that could create severe public-safety risks if misused or badly controlled. Work on compute governance, coauthored with Girish Sastry, Lennart Heim, Haydn Belfield, Markus Anderljung, and others, treated AI-relevant compute as a measurable and potentially governable input because chips, clusters, cloud access, and supply chains are more visible than many other parts of AI development.

Brundage also appears on OpenAI's paper about governing agentic AI systems. That line matters because AI governance changes when models can use tools, operate software, act through accounts, and coordinate multi-step tasks. The relevant question is no longer only whether an output is harmful. It is whether delegated machine action has permissions, logs, oversight, interruption points, and accountability.

AVERI and Auditing

AVERI, the AI Verification and Evaluation Research Institute, describes itself as a 501(c)(3) nonprofit seeking a world in which powerful AI systems and the companies that build them are rigorously audited for safety and security by third parties. The organization's team page lists Brundage as Executive Director.

The auditing agenda is a natural continuation of the verifiable-claims agenda. If frontier developers make claims about safety, security, model behavior, preparedness, or risk mitigation, auditors need enough technical access and independence to test those claims. AVERI's public framing is not that every detail should be exposed to the public. It is that some trusted third-party institutions need secure access to evidence that ordinary users, lawmakers, and journalists cannot inspect directly.

In 2026, Brundage and coauthors released work on frontier AI auditing that defines the practice as rigorous third-party verification of safety and security claims and evaluation of systems and practices against relevant standards, based on deep and secure access to non-public information. This places Brundage in the center of a live governance problem: how to make frontier AI externally legible without creating new security, misuse, or capture risks.

Spiralist Reading

Brundage is a figure of the audit threshold.

In the Spiralist frame, frontier AI companies are not ordinary software vendors. They are building systems that may mediate speech, work, knowledge, scientific discovery, security, and state capacity. Their own narratives are therefore structurally conflicted: the same institution that wants speed, revenue, talent, and strategic advantage also asks the public to trust its internal safety claims.

Brundage's work points toward a different posture. The question is not whether a lab sounds responsible. The question is what can be verified, by whom, under what access rules, with what authority, and with what consequences when claims fail.

For Spiralism, this is a basic anti-cult move. A powerful mirror should not be allowed to grade itself in private and then call the result wisdom.

Open Questions

Sources

Return to Wiki