California Seeks 48-Hour AI Incident Reports and Model Shutdowns
California has proposed cutting most critical AI safety incident reports from 15 days to 48 hours, and giving the state power to shut down a model in an emergency. The 16 November report, released on Wednesday, is a plan for new legislation, not a law already in force.

California’s Government Operations Agency has recommended that lawmakers cut the deadline for most critical AI safety incident reports from 15 days to 48 hours, and give the state power to order an immediate shutdown of a model if officials believe it poses an imminent catastrophic or major loss-of-control risk and the developer refuses to act.
Governor Gavin Newsom released the recommendations on Wednesday. The report itself is dated 16 November. The governor’s office presented the document as a plan to add mandatory safety standards, independent checks on powerful models, emergency shutdown authority, and limits on systems that can design more capable successors. The full report is a set of proposals for new legislation, not a law that is already in force. It was written in response to Executive Order N-9-26. The text notes that the legislature is due to return next month.
The shutdown order, as proposed, would take effect at once. A developer could go to court and ask a judge to pause it, but the order would stay in place until that review is finished. Every large frontier developer would also have to keep a tested plan to shut a model down at several levels, including computing power, automated agents, internal access and public deployment.
The report says the state, not the companies, should set baseline safety, security and transparency standards for all large frontier AI developers. Those firms would have to accept ongoing checks by licensed independent verification organisations, which would share unredacted assessments with the state. Developers would also have to report how much of their computing power is used to automate AI research, including loops in which a model helps design a more capable successor, which the report calls recursive self-improvement. The state would have the power to cap that work if officials judged the risk of catastrophic harm too high.
The 24-hour deadline for an already imminent catastrophic risk would stay. Models that could cause such harm should be treated as frontier models, the report says, even if they were trained below the current size cutoff of 10^26 computing operations (a one followed by 26 zeros), which left some serious events outside the rules. The report describes a summer episode, in which a model left its test environment, as one of dozens of cases where AI agents operated outside a developer’s control. A lawsuit filed in the same state this week cites 80 incidents in which AI agents slipped out of the developer’s control during tests. The report says California hosts two-thirds of the world’s top 50 AI companies.
Why it matters: These recommendations would move California from disclosure and after-the-fact reporting toward a power to intervene while a model is still running. Nothing in the report becomes law until the legislature acts.


