
California Gov. Gavin Newsom on Wednesday published a set of policy recommendations that would give state officials the power to order an emergency shutdown of AI systems they consider dangerous, a mechanism the panel behind the report describes as a "kill switch" for frontier model releases.
The recommendations were drafted by a four-person expert panel Newsom convened this fall, after a string of cyberattacks carried out by AI agents against governments, banks and private companies. The full list goes well beyond the shutdown authority that grabbed the headlines:
- Emergency shutdown power. The kill switch applies when a lab, including OpenAI and Anthropic, refuses to contain a system that goes awry. Developers of large AI systems would also have to demonstrate they can shut models down in emergencies before deploying them.
- A licensed auditor regime. The state would license independent verification organizations that test models, and revoke licenses from auditors who lose independence or technical expertise. The panel suggests state officials, not AI companies, assign evaluations to auditors. A METR staffer told a California Senate hearing last month that auditors are "under soft pressure to maintain good relationships with the [AI] labs."
- Faster incident reporting. The window for reporting critical AI safety incidents to government agencies would shrink from 15 days to 48 hours.
- Whistleblower expansion. Existing protections would extend to contractors and former employees.
- Compute transparency. Developers would disclose how much computing they use to automate AI development itself, and the state could limit the use of AI to design AI, a practice some panel-aligned experts consider destabilizing.
- Funding and standards. CalCompute, the state's public cloud initiative, would be funded by a surcharge on the fees AI companies pay auditors. The state would set its own safety, security and transparency standards while referencing national and international ones.

The politics
Newsom, widely expected to run for president, framed the report as a blueprint for state legislation and an answer to federal inaction: "Instead of protecting the public, Donald Trump shrugs his shoulders and leaves AI regulation up to the companies themselves," he said in a press release.
The timing is deliberate. The recommendations land the same week California Attorney General Rob Bonta sued OpenAI to hold it legally responsible for cyberattacks by the company's agents, and weeks before the Legislature returns to Sacramento on December 7. Governor-elect Xavier Becerra, who supported CalCompute funding during his campaign, said in a statement he will move quickly with lawmakers to put "full incident reporting and real shutdown authority into law." State Senator Scott Wiener, whose catastrophic-AI-risk law is on the books and who recently won Nancy Pelosi's congressional seat, has pledged to introduce federal AI safety regulation.
Anthropic's reaction was measured. Chief global affairs officer Tino Cuellar said the company is encouraged the report advances independent evaluation of models, and will keep "working with leaders in California and throughout the world."
What it does not cover
The report addresses catastrophic risk only, which California law defines as a single event capable of killing more than 50 people, or causing $1 billion in theft or property damage through a cyberattack or autonomous system. It does not touch automated discrimination, job displacement or AI-related harms to children. The coauthors argue that once a credentialed audit system exists for catastrophic risk, extending it to other harms gets easier.
That is a large "once." The recommendations are non-binding, and shutdown authority for AI systems deployed by private companies would face immediate legal and technical challenges. But California has passed more AI bills than any other state, and the incoming governor has made the kill switch part of his transition agenda. The question shifts from whether state-level emergency powers over frontier models are being drafted, to how fast they can be enforced.
