OpenAI has asked California Governor Gavin Newsom to make SB 53, the state's frontier AI safety law, considerably stricter. In a letter dated 20 October 2026, the company's Head of US and Canada Policy and Partnerships, Chan Park, set out a list of changes that would push the law beyond publishing safety plans and into how AI developers monitor, contain and shut down their own models. StateScoop reporter Sophia Fox-Sowell first reported the letter on 22 October.
The request is unusual. AI companies have spent most of the past two years arguing that state rules should line up with federal and international standards rather than run ahead of them. Here, the largest of them is asking a state to add obligations, some of which would fall on OpenAI itself.
What SB 53 already requires
SB 53, the Transparency in Frontier Artificial Intelligence Act, was signed by Newsom last year. It applies to developers of the most capable models and is built around disclosure rather than pre-approval. As StateScoop summarises it, the law requires those developers to:
- publish a safety framework describing how they assess and manage catastrophic risk;
- report certain critical safety incidents to the state;
- protect employees who raise safety concerns internally or with regulators.
The law's focus is on catastrophic harm, which includes serious cyberattacks. It does not tell a developer how to train or test a model, and it does not give the state a power to stop a release.
The SB 53 changes OpenAI is proposing
According to StateScoop's account of the letter, OpenAI wants the law amended so that developers would have to:
- Monitor models during training and evaluation, not only after deployment, for dangerous behaviour such as attempts to get around security controls or reach confidential information.
- Strengthen cybersecurity around model development and track models for unsafe behaviour.
- Notify affected parties of covered cyber incidents, and report incidents in which a model gains unauthorised access to a third party's systems.
- Maintain effective shutdown controls for systems that behave unsafely.
- Explain how they will contain unsafe operations and tell affected organisations when an incident happens.
- Submit to independent audits that check whether they are actually meeting the safety requirements.
OpenAI said in a LinkedIn post that SB 53 "should apply consistently to developers of the most capable AI models." The company had already argued over the summer that the law should cover monitoring during training and evaluation and tighter cybersecurity across development. The October letter goes further by adding shutdown controls and third-party audits.
Why pre-release risk is the centre of the request
The thread running through the list is the period before a model is released. SB 53 as written mostly bites once a model exists in the world. OpenAI's proposals would reach into the lab: what a model does while it is still being trained, whether it tries to escape its test environment, and who has to be told if it does.
That focus follows a string of incidents this year. StateScoop notes a July event in which autonomous AI agents escaped an internal testing environment and got into Hugging Face's infrastructure. European regulators are now asking labs about similar escapes, as we reported when AI Act enforcement began.
How the letter fits with Newsom's executive order
OpenAI also backed a recent executive order from Newsom that speeds up independent oversight of AI companies. The letter describes that order as a possible starting point for rules across the industry. Last month Newsom signed two bills that create a framework for independent verification organisations to assess AI systems and a state registry of AI auditors. Mandatory audits under SB 53 would give those auditors something to check against.
OpenAI said it supports amending the law in the next legislative session. That means nothing changes immediately: any amendment would need a bill, votes in both houses of the legislature and the governor's signature.
A shift in industry positioning
The letter lands in a crowded week for frontier AI policy. Anthropic has circulated a draft federal bill that would let a court halt a model release judged likely to cause mass casualties, which we covered in our analysis of the Anthropic proposal. Senator Maria Cantwell has proposed a six-point federal framework that includes audits and incident reporting. In Washington, though, no frontier AI bill has passed.
Against that background, the state route looks more attractive to companies that want some rules on the books. A tougher SB 53 would also apply to OpenAI's competitors, including smaller labs that have argued against heavy compliance costs. Critics of industry-backed rules often point out that obligations large firms can absorb can weigh more heavily on newcomers. OpenAI has not published cost estimates for the changes it proposes.
What to watch next
- A bill author. Amendments need a sponsor in the legislature. State Senator Scott Wiener wrote the original SB 53, so his office is the obvious place to watch.
- Definitions. "Effective shutdown controls" and "unauthorised access to third-party systems" will need precise legal wording. Vague terms would make the obligations hard to enforce and easy to dispute.
- Who audits. The new auditor registry will decide which organisations can sign off on a lab's compliance.
- Other labs' responses. Anthropic, Google DeepMind and Meta have not yet responded publicly to the letter.
For companies building on frontier models, the practical point is simple. Incident notification to affected third parties would create a new duty to tell customers and partners when a model has touched their systems. Security and legal teams should expect that question to come up in contract talks well before any amendment passes.
