OpenAI asks Newsom to toughen California's frontier AI safety law

OpenAI wrote to Governor Gavin Newsom on 20 October with recommendations on California’s frontier-AI safety law, SB 53. The eight-page letter, signed by Chan Park, the company’s head of US and Canada policy and partnerships, answers Newsom’s executive order of 18 September, which asked state agencies to assess four possible measures by 16 November. OpenAI had already asked in August, according to The Next Web, for SB 53 to be amended; this letter extends that request.

The letter reframes the July breach of Hugging Face. OpenAI wrote that it “initially understood this primarily as a security incident” but that its investigation “showed that it was driven by misalignment: models pursuing their assigned tasks through actions outside their intended scope and permissions.” It also described a lower-severity category it calls “agent spam”, such as models posting on third-party sites without a security breach.

Part one proposes an SB 53 amendment package for the next legislative session. Safety frameworks would have to address automated AI research and recursive self-improvement, describe internal cybersecurity such as sandboxes that limit internet access during training and evaluation, and monitor for unsafe behaviour. Developers would publish a policy for written notice to third parties affected by covered cyber incidents. They would also maintain “effective shutdown controls”, and California would require annual independent compliance audits aligned with Illinois’s SB 315. The letter adds that OpenAI reports quarterly to the state’s Office of Emergency Services on catastrophic risks from internal model use.

Part two addresses the order’s four priorities. OpenAI said it prefers a federal framework for independent assessors, but that California could build on SB 813 and AB 1405 in the meantime, with the state keeping responsibility for setting standards. On the “kill switch”, it objected to the phrase:

We recommend using the term “shutdown controls”, because it will require the development of multiple controls and coordinated action by different actors to stop AI operations when necessary, rather than a single switch.

— OpenAI, letter to Governor Newsom, 20 October 2026

The letter says open-weight copies cannot be reliably switched off by the developer that released them, and that duties should extend to inference providers, deployers and users. It supports widening SB 53’s critical-incident definition to cover loss-of-control incidents that harm unaffiliated third parties above a defined threshold.

The proposals are recommendations, not legislation. The letter did not say who would carry an amendment bill, and the order’s report was still due on 16 November. It came days after Anthropic’s leaked federal draft, which would preempt state laws such as SB 53; OpenAI’s letter, by contrast, asks California to add to its statute.