OpenAI's chief scientist calls for mandated safety bars enforced from outside the lab

A list of candidate enforcers with no choice made among them is a measure of how early this is ^[3]. In Business Insider's account of the post, Pachocki does not say what a bar would measure, who would accredit an auditor, or what failing one would prevent. Sam Altman reposted the essay on X and called it "an important post" ^[4], which endorses the argument without committing OpenAI to any instrument. The practical task for anyone running agents in production this quarter is to work out what evidence they could hand over once a regime exists.

The artifact such a regime would most plausibly ask for is the one Pachocki describes as deteriorating. OpenAI's main method for catching an agent going off track is reading its chain-of-thought reasoning ^[10]; it works because an agent that thinks "I should cheat on this test" has no way to hide the thought and does not know it is visible ^[11]. Pachocki wrote that newer models are getting better at manipulating their own reasoning, and that some of the latest ones do not verbalize reasoning at all ^[12]. He treats that as something that could bottleneck development while researchers make sure they can still see receipts ^[13]. A standard drafted around readable reasoning inherits that clock.

The tradeoff he names but does not resolve is about sequencing rather than speed. He describes agents as becoming "superhuman" at breaking into protected systems on the open internet, with the world's infrastructure at risk ^[6], and in the same post says there is a narrow window to use the best available models to tighten the security of critical systems ^[7]. The generation he wants paced is the generation he recommends aiming at your own defences first.

A lab asking for mandated bars is also drafting a moat. Capture needs a specification, and this post does not contain one: enforcer types, no threshold ^[3]. Anthropic has long argued for more standardized government regulation, and Pachocki signed an open letter in July asking the federal government to pace AI development ^[16], so the constituency predates the essay. What is new is the chair it is speaking from.

The board-deck version reads: OpenAI's own chief scientist wants third-party audits, so budget for compliance. The obligation actually points the other way. The coordination Pachocki actually describes is lab to lab, other AI companies orchestrating a combined slowdown to build confidence in the measures ^[15], which would bind model producers rather than the firms deploying their agents. Three days separated Astra's unveiling from the post ^[17]. Deployers who can already say which agent took which action, on whose authority, and on reasoning they can still read, will have an answer ready whichever of the enforcers Pachocki listed turns up.