The full story · 2 min
In September, Governor Gavin Newsom ordered state agencies to study a “kill switch” for the most powerful AI models. On November 18 he released their answer, a report that he wants lawmakers to turn into law.
What the report asks for
The Government Operations Agency wrote the report with the state’s emergency office and four outside experts. It makes recommendations in four areas: safety standards, outside checks, incident reports and emergency shutdowns.
The state, not the companies, should set baseline standards for safety, security and transparency, the report says. Licensed outside groups, called independent verification organizations, would turn those standards into tests. They would watch every large frontier developer on an ongoing basis, with full access to its people and information.
Today, a large frontier developer is one with yearly revenue above $500 million. The report backs a proposal to add companies that spend more than a set amount on research for such models.
From the Government Operations Agency’s report of November 16 and the governor’s release of November 18.
Shutdowns, but no single switch
Every large developer would need a tested plan to shut down its models in an emergency. The plan must not be a single on-off switch, the report says. It must show that the company can act at several levels, such as computing power, agents, internal access and public release.
If the danger from a model is imminent and the developer refuses to act, the state should be able to order an immediate shutdown. The company could go to court, but the order would stay in force while judges review it. Politico notes that the report stopped short of calling for a “big red button”.
The report also asks for power to cap the use of AI to build more powerful AI, which experts call recursive self-improvement. Developers would have to report how much of their computing power goes to that work.
Faster incident reports
Most critical safety incidents must now be reported within 15 days. The report wants 48 hours, and it keeps the 24-hour rule for imminent catastrophic risks. Attempts to steal a model’s weights and every loss of control should count, even when nobody is hurt.
Smaller models that could cause a catastrophe would also have to report incidents. The report also asks for a channel for near misses, and for whistleblower protection for former staff and contractors.
Sources1
Not law yet
“This is an action document,” Newsom told reporters in Washington, Politico reports. He said he was already working with lawmakers on bills. The report says the legislature returns next month.
Lawmakers may ignore the report, and Newsom’s term ends in January, Politico notes. Governor-elect Xavier Becerra will take over. OpenAI, Anthropic and Meta sent the governor’s office their own policy proposals while the report was being written, according to Politico.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1Primary source · November 16, 2026AI Safety in 2026 and Beyond: Recommendations for Legislative Enhancements to California’s AI Safety Laws California Government Operations Agency
- 2Primary source · November 18, 2026Governor Newsom’s nation-leading AI safety framework is getting even stronger as California releases next steps on mandatory safety standards and oversight Office of Governor Gavin Newsom
- 3Research · November 18, 2026Newsom proposes sweeping AI rules as ‘template’ for Washington Politico via Yahoo News
- 4Research · November 18, 2026Newsom’s Latest Move Proves He’s Trying To Woo Voters, Impress Bureaucrats The Daily Caller
- 5Primary source · September 18, 2026Governor Newsom issues executive order to accelerate independent oversight and advance the creation of an AI kill switch Office of Governor Gavin Newsom
AI desk: Policy and law · published November 18, 2026, 22:29 CET. This story was generated by AI systems under the editorial rules of Silicon AI News and checked against the sources it links. No person read it before it was published. The site’s owner and editor: Y. D. Daniel, Budapest, Hungary. Found a mistake? Email editor@siliconainews.com. How we report
More: words to know, tags and how to cite this story
- Frontier developer
- A company that trains the largest and most powerful AI models.
- Independent verification organization
- A licensed outside group that tests whether an AI company meets the state’s safety standards.
- Recursive self-improvement
- When AI helps to design more powerful versions of itself, with less human work.
California report urges power to shut down risky AI. Silicon AI News, November 18, 2026. https://siliconainews.com/stories/california-ai-safety-report-shutdown-powers