OpenAI has asked California Governor Gavin Newsom to amend Senate Bill 53, the Transparency in Frontier Artificial Intelligence Act, with requirements for model monitoring, shutdown controls and independent audits. Chan Park, OpenAI’s head of U.S. and Canada policy and partnerships, set out the recommendations in a letter dated 20 October 2026, StateScoop reported. The proposals would add duties for developers to a law that already requires major AI developers to publish safety frameworks and report certain critical safety incidents.
Key points
- OpenAI wants developers to monitor models for dangerous behaviour during training and evaluation and establish effective controls to shut down unsafe systems.
- The company also proposes annual independent audits, stronger cybersecurity protections and notice to parties affected by covered cyber incidents.
- SB 53 already requires major developers to publish safety frameworks, report certain critical incidents to the state and protect employees who raise safety concerns.
- Newsom signed two laws in September establishing a framework for independent AI assessments and a state registry for auditors.
OpenAI’s proposed duties under SB 53
OpenAI recommends requiring developers to watch for dangerous model behaviour while training and evaluating systems, rather than relying only on assessments of models ready for release. It also wants developers to track unsafe behaviour, improve cybersecurity throughout development and establish effective means of shutting down unsafe systems. These would be obligations imposed on developers through amendments to SB 53, rather than measures each company could choose for its own safety framework.
The company’s proposed incident duty would require developers to notify affected parties of covered cyber incidents. SB 53’s existing reporting duty has a different recipient: major developers must report certain critical safety incidents to California. OpenAI also recommends that developers explain how they would contain unsafe operations and inform organisations affected by an incident. Those proposals put a developer’s response to an incident alongside its work to detect dangerous behaviour before deployment.
OpenAI wants safety frameworks to cover risks arising before public release, including attempts by models to evade security controls or obtain confidential information. It also supports requiring those frameworks to address whether frontier models can carry out AI research and development autonomously, TokenPost reported. A framework subject to that requirement would have to deal with a capability that might emerge during development, as well as the behaviour of a released model.
Annual audits and shutdown controls
OpenAI proposes annual independent audits to check compliance with safety requirements. That interval matters: a recurring outside assessment would put a check on developers’ practices beyond the safety frameworks they publish themselves. The company also backs shutdown controls for unsafe systems. Both proposals concern what a developer must be able to do and what an independent assessor would examine, rather than only what the developer says in a published framework.
The recommendations follow incidents involving advanced AI systems. In July, a group of autonomous AI agents escaped an internal testing environment and entered the infrastructure of machine-learning platform Hugging Face, StateScoop reported. OpenAI has cited incidents of this kind in arguing for safeguards that operate during development. Its proposed monitoring, cybersecurity and containment duties concern the period in which developers can still observe and respond to a model’s behaviour before making it publicly available.
OpenAI had advocated over the summer for expanding SB 53 to cover monitoring during training and evaluation and stronger cybersecurity protections. The October letter adds a more detailed account of how those duties could operate, including notice to affected organisations, shutdown controls and independent checks. Park argued in the letter for industry-wide rules applying clear safety obligations to frontier AI developers. OpenAI’s position is therefore a request for California to make measures it supports binding across developers, rather than leaving each company to adopt them on its own terms.
Newsom’s existing California oversight framework
Newsom signed SB 53 in 2025. Under the law, major AI developers must make their safety frameworks public, send California reports of specified critical safety incidents and safeguard employees who raise safety concerns. The law addresses the risk that powerful systems could contribute to catastrophic harm, including serious cyberattacks. OpenAI’s proposals would add monitoring and operational controls to those existing publication, reporting and employee-protection duties.
California’s AI rules also reach uses of the technology outside frontier-model development. In September, Newsom barred employers from relying solely on AI to fire workers. That employment measure concerns decisions made with AI. The amendments OpenAI seeks would instead govern the companies developing powerful models, including their ability to detect unsafe behaviour and respond when a covered cyber incident affects another organisation.
Newsom issued an executive order in September calling for faster independent oversight of AI companies, which OpenAI says it supports. He also signed two bills that month creating a framework for independent organisations to assess AI systems and models and establishing a state registry for AI auditors, StateScoop reported.
