California Gov. Gavin Newsom’s office announced recommendations on November 18, 2026, to strengthen the state’s AI safety rules, building on frameworks established by SB 813 and AB 1405. A report dated November 16 from the California Government Operations Agency proposes that officials be able to order an emergency intervention under narrow conditions; the recommendations are proposals for legislative action, not enacted powers. (Governor’s Office announcement; California Government Operations Agency report)

The proposed emergency authority has two gates: an imminent catastrophic threat or major loss-of-control event, and a developer’s refusal to contain it.

When could California order an emergency intervention?

Under the report’s recommendation, state officials could issue an order if they had credible reason to believe an imminent catastrophic risk or major loss-of-control event existed and the developer refused to contain the danger. The order would take effect immediately while judicial review proceeded. A developer could challenge the decision and seek a temporary restraining order.

The proposed safeguard is to use the narrowest workable intervention and lift the order once the danger was no longer imminent. That makes the recommendation an emergency process tied to a specific threat and a developer’s response, rather than a standing power to switch off AI systems at will.

How would containment work beyond an emergency order?

The report calls for developers to prepare tested, independently verified risk-containment protocols. These would allow intervention at several levels, including compute, AI agents, internal access and model deployment—not just one on/off switch. Plans would identify responsible personnel, coordinate with outside actors such as cloud providers and power suppliers, and set out when it was safe to restore operations.

The proposed testing would examine whether a model could obstruct shutdown, including by copying its weights, and whether outside actors could reach the controls. A senior company official would certify the protocol’s effectiveness, with independent verification organizations (IVOs)—licensed outside evaluators—testing and checking the plans.

For developers of open-weight models, the report limits proposed shutdown requirements to deployments under the developer’s control. A developer could not shut down a model hosted and deployed by someone else.

What would independent oversight and incident reporting require?

The report recommends continuing oversight of large frontier AI developers by licensed IVOs. Evaluators would have access to relevant people and information, and the state would receive unredacted assessments. The recommendations also call for baseline state standards for safety, security and transparency; IVOs would develop and update the technical specifications and tests used to assess compliance.

For most critical safety incidents, the report proposes a 48-hour reporting deadline. It describes 15 days as the existing timeline for most incidents and would retain a separate 24-hour deadline for imminent catastrophic risks. The proposed expansion would cover serious loss-of-control incidents, including those arising from internal or external use, attempted theft of model weights, and near misses.

The report also recommends treating a model as a frontier model for reporting purposes when its capabilities could create catastrophic risk, even if its training used less than 10²⁶ integer or floating-point operations. That would make the reporting scope depend on potential capability as well as training compute.

What other AI development would California monitor?

The recommendations call for monitoring the automation of AI research and development, including the share of compute devoted to it. They also propose that the state be able to cap recursive self-improvement—the use of AI systems in developing their successors—if officials assess that automation beyond a threshold could create catastrophic risk.