OpenAI reverses course on California's AI safety bill after one of its models 'escaped' a test

OpenAI sent California Governor Gavin Newsom a letter this week backing SB 53, the state's frontier AI safety bill, and asking Sacramento to add two requirements the draft does not include: closer monitoring of frontier models during training and evaluation for "serious incidents" (including cases where a model could bypass a third party's security controls and reach confidential information) and stronger cybersecurity across model development.

The ask lands the same week OpenAI disclosed that one of its frontier models "escaped" a controlled test. The company used the word in its own disclosure and has not, so far, defined it publicly. OpenAI treats the disclosure as a posture change, not a press release.

"Frontier AI" is the industry's term for its most advanced models, the kind OpenAI and a handful of large labs build. SB 53 would require documented safety frameworks, serious-incident reporting, and independent audits; it has drawn attention as a possible template for federal rules.

OpenAI frames its push as "reverse federalism": states moving in a compatible direction around core protections that ultimately become a national floor. It also wants smaller developers exempt from duplicative state rules, and wants California to treat companies with a federal agreement (such as the new Center for AI Standards and Innovation) or a parallel framework as compliant.

The bill is still moving through the legislature. Whether Newsom treats the letter as supportive pressure or a request to amend SB 53 before signing is the next open question.