FULL STORY

Ex-OpenAI Policy Chief Brundage Bets on Frontier AI Audits

Days after leaving OpenAI, Miles Brundage announced he would dedicate himself to frontier AI auditing, then expanded his views in a series of posts arguing AI safety is a sociotechnical problem without an endpoint.

2026-09-16 ~ 2026-09-18 · 2 episodes · 18 posts

Episode 1 · Ex-OpenAI policy chief Brundage pivots to frontier AI auditing (2026-09-16, 9 posts)

Miles Brundage, former OpenAI policy lead, announced on Sept 16 that after leaving the company he will devote himself fully to "frontier AI auditing," arguing third-party audits are worth voluntary industry adoption—while stressing audits are only one necessary condition and can easily become box-checking exercises. He laid out a full slate of governance proposals and answered critics who asked why he didn't call for a pause or slowdown.

Confirmed

  • Brundage argued audits should be mandatory at top AI companies, with minimum standards for auditor qualifications, independence, and access; real safety standards are needed rather than compliance via "having a (possibly weak) policy and following it."
  • On penalties, he noted that today "the cost of breaking the law is less than one senior engineer's salary," and violations must carry truly serious consequences; he also called for much greater investment in biosecurity, cyberdefense, alignment research, and safety systems resilient to AI-enabled nation-state attackers.
  • He advocated strong whistleblower protections, saying NDAs currently deter people from telling policymakers about "legal but dangerous" practices. On Sept 17 he shared a Lawfare piece, "Protecting AI Whistleblowers," calling it "the easiest no-brainer" that lets employees inform government and builds knowledge for every future AI bill.
  • He proposed mandatory AI incident sharing to inform the public and support research and policy, with transparency requirements extending to full disclosure of behavioral specs/constitutions.
  • Responding to "why no pause/slowdown" criticism, he said auditing must be done well first—high standards, serious penalties, mechanisms that make violations visible—before a pause or slowdown could actually work; other governance work is needed regardless, and once auditing works, necessary slowing would follow naturally.
  • He said he feels "okay" for now because broad recognition that the industry lacks adequate control over frontier AI is growing—but this is only a start: "lots to do, and not much time."

Why it matters

  • Having sat at the core of OpenAI's policy and safety apparatus, Brundage's pivot to independent auditing work is a direct response from a former top-lab executive to the inadequacy of current governance, and may shape industry and regulatory agendas.
  • By positioning auditing as a precondition for more aggressive measures like pauses, complemented by whistleblower protection and incident disclosure, he offers an actionable middle path in the debate over how to govern frontier AI.

Episode 2 · Ex-OpenAI policy chief Brundage: AI safety is sociotechnical; security should top AI policy (2026-09-17, 9 posts)

Miles Brundage, former OpenAI policy lead, published a series of long posts on 09-17 and a follow-up essay on 09-18 laying out his view of AI safety and policy priorities: AI safety is not a purely technical problem but a sociotechnical one with no one-shot solution; he argued that security—preventing AI technology from being stolen or tampered with—should be the top AI policy priority, and that security progress failing to keep pace with capability gains is the strongest argument for slowing down.

Confirmed

  • Brundage argued that anti-regulation voices framing AI safety as a "purely technical problem" is a motte-and-bailey move: it makes the problem seem easy to solve and shields the industry from criticism.
  • He analogized AI safety to complex sociotechnical systems like aviation, nuclear power, and healthcare: without constraints, under time pressure and competition, people cut corners, slack off, and fall into groupthink—technical fixes alone cannot solve this.
  • He highlighted a "safety gains reinvestment" effect: safety headroom gets converted into larger pretraining runs, harder RL environments, more test-time compute, and looser internet access; there is no clear ceiling on AI capability, and no company will declare "we're safe enough" and stop.
  • Discrete safety improvements are insufficient—gains are reinvested into new risks; even if a recipe for perfectly safe/aligned frontier models existed tomorrow (he stressed it does not), companies would still cut corners chasing the next tier under competition.
  • Drawing on nuclear power: even with extensive design-specific safety regimes, there has never been a universal "solve it once" fix—AI likewise requires continuous iteration.
  • He proposed a test: any AI governance theory that cannot survive historical backtesting against dangerous systems is not credible; a single bad act is not doomsday, but without guardrails it can escalate.
  • On 09-18 he wrote that security against theft and tampering should be the top AI policy priority, explained his shift toward the issue, said he would work on it if he had more to contribute, and noted with regret that security progress lags capability growth—perhaps the strongest argument for slowing down.

Why it matters

  • This is a systematic argument from a former OpenAI policy chief, directly rebutting the "AI safety is merely technical, no regulation needed" narrative and offering a conceptual framework for AI governance debates.
  • The "safety gains reinvestment" and "no capability ceiling" arguments imply safety problems accumulate as capabilities scale, with direct implications for policymaking and corporate governance.
  • His elevation of anti-theft, anti-tampering security to top policy priority, and his frank admission that security lags capability, give the "slow down" camp an argument from a former insider at a leading lab.