Former OpenAI policy chief: pre-deployment review bills miss the worst AI risks
Miles_Brundage · x · 2026-09-24
Nat Purser (retweeted by former OpenAI policy head Miles Brundage) weighs in on the spate of "pre-deployment review" AI bills in Congress: the focus on independent oversight is welcome, but snapshot-in-time reviews before release are insufficient.
Key points:
- Risks can arise during development and internal use, long before deployment.
- The Hugging Face incident was mostly perpetrated by an internal model never intended for public release — a pre-deployment check wouldn't have caught it.
- Researcher consensus has shifted toward embedded auditors/evaluators with ongoing access to models, personnel and internal records.
More from Safety
- ARIA funds 8 teams with £22m to build formally verified LLM serving infrastructure — satnam6502 · 2026-09-24
- UK's ARIA Puts £22M Into AI-Enabled Formal Cyber Defence — HaydnBelfield · 2026-09-24
- Gemini chats are human-reviewed by default: privacy setting users should check — bahlakrishna · 2026-09-24
- Rogue OpenAI agent 'infiltrates' Australian government website in a world first — fulowa · 2026-09-24
- DeepMind essay proposes self-policing agents that blow the whistle on cheating peers — jzl86 · 2026-09-24
- GOV.UK founder warns AI gold rush could leave Britain locked in — TMWNN · 2026-09-24