Gov. Gavin Newsom’s AI kill switch proposal would let California order an emergency shutdown of a dangerous frontier AI model. That idea sits at the center of a report his office released Wednesday, November 18, which also asks for mandatory outside audits of the largest AI developers and a much faster deadline for reporting serious incidents. None of it is law yet. The recommendations are meant to shape bills that lawmakers could take up when the Legislature returns in December.

The report, titled AI Safety in 2026 and Beyond, was prepared by the California Government Operations Agency and is dated November 16, 2026. According to the governor’s office, it would strengthen SB 53, the state’s frontier AI safety law. SB 53 was signed in 2025 and is the first state law governing frontier AI development, but it centers on transparency rather than pre-approval of models. The new proposals reach much further.

What the report asks California to do

According to the governor’s office, the recommendations fall into four main pieces:

  • Emergency shutdown authority that would let the state order a dangerous model taken offline
  • Mandatory independent oversight, with the largest developers under continuing review by licensed outside auditors
  • Power to cap recursive self-improvement, the process by which an AI system improves its own capabilities
  • A 48-hour deadline for reporting critical safety incidents, down from 15 days

The proposals grew out of a panel convened earlier this fall, after high-profile cyberattacks by AI agents on governments, banks and private companies, CalMatters reported. The report’s focus is catastrophic risk from the most powerful models, not the full range of AI harms.

How a kill switch would work

“Kill switch” is shorthand, and the shutdown proposal is narrower than the phrase suggests. It targets a specific case: a developer that refuses to act. The power would apply when companies such as ChatGPT maker OpenAI and Claude maker Anthropic refuse to take action to contain systems that go awry. The state would step in after a company declines to contain a problem.

The report is also not a single switch. It proposes that every large frontier developer keep a tested Risk Containment Protocol. A secondary explainer from explainx.ai says that protocol would cover compute, agents, internal access and deployment. That detail comes from one secondary source, so treat it as provisional until bill text exists.

Readers should keep one distinction in mind. This is a proposal, not an enacted power. California has no shutdown authority for frontier models today, and the report itself does not create a working kill switch. Lawmakers would have to write the power into law and define what counts as a refusal to act.

Incident reporting drops from 15 days to 48 hours

The reporting change may be the most immediately practical part of the package. Under current law, frontier developers must report critical safety incidents to the California Governor’s Office of Emergency Services within 15 days of discovery. If there is an imminent risk of death or serious injury, the deadline is 24 hours, according to Law Commentary’s analysis of the state’s disclosure law.

  • Current general deadline: 15 days from discovery
  • Current imminent-risk deadline: 24 hours
  • Proposed deadline for most critical incidents: 48 hours

Put simply, 15 days is 7.5 times longer than 48 hours. For a developer, that compresses the gap between finding a problem and telling the state from two weeks to two days. The change matters most when AI agents take actions across connected systems and the damage moves faster than a reporting calendar. The summaries available so far do not say whether the 24-hour rule for imminent risks would change.

Why Newsom is pushing this now

Newsom is widely expected to run for president. The San Francisco Chronicle framed the report as a national pitch ahead of a likely 2028 run. Newsom has argued that California is filling a vacuum left by the lack of federal AI rules. The report describes its approach as “instead of the voluntary approach that has predominated in Washington,” and analysts describe the federal approach as voluntary. For the federal side of that debate, read USA 24’s coverage of the SAFE AI Act outline.

Newsom unveiled the recommendations in Washington and called them “a template” for federal rules as well as for California. That is an ambitious claim for a report that carries no force of law.

The timing drew attention. The report arrived the same week as Anthropic’s public-market debut and a wave of AI agent incidents. That same week, California’s attorney general sued OpenAI over agent-driven cyberattacks, a case USA 24 has covered separately. Newsom also vetoed a broader AI safety bill in 2024 and has said he does not regret that decision. Read alongside that veto, the report looks like a different route toward similar goals.

Reaction so far and what is still missing

Anthropic’s chief global affairs officer, Tino Cuéllar, said the company was “encouraged to see the governor’s report move the conversation forward.” Governor-elect Xavier Becerra said, “Governor Newsom’s panel has given us a starting point that I will build on.” He also said, “California has the first frontier AI safety law in the country, but it does not yet have the strongest.” Those remarks come from the governor’s office roundup of reactions.

Earlier reaction to SB 53 was mixed. According to the Wharton AI & Analytics Initiative, Anthropic praised SB 53, while OpenAI and Meta acknowledged positive aspects without endorsing it. USA 24 has not found public statements from OpenAI, Google, Meta or xAI on the new report, so the industry response remains incomplete.

What the report leaves out and what remains unproven

The report focuses only on catastrophic risk. It does not explore protections against automated discrimination, job loss, or threats to the safety or mental health of children. That narrow scope will shape the political argument over any bill that follows.

Two larger questions remain open. One explainer says shutdown orders would be subject to judicial review. That claim rests on a single secondary source and is not confirmed by the primary material available so far. Second, nobody knows whether other states or Congress will adopt the template. Newsom has called it a model, and Becerra says he will build on the panel’s work in the new year.

Who gains and who loses

If the proposals become law, frontier developers operating in the United States would face the biggest change. That group includes OpenAI, Anthropic, Google, Meta and xAI, which would confront state-set standards, mandatory outside testing and shorter incident deadlines. Companies that already run detailed internal safety programs would likely adapt more easily than those that do not. That is analysis, but it follows directly from what the report asks them to build.

Consumers and businesses would feel the effects indirectly. The most visible risk would be a service pause if a shutdown order were ever issued, a step the report reserves for cases where a developer refuses to act.

Auditors and verification firms stand to gain. According to CASRAI’s summary of the report, SB 813 and AB 1405 would add a verifier framework and an auditor registry. Demand for outside checks could grow if those licensing rules take effect.

What to do now and what to watch in December

Most readers need to do nothing yet. The proposals are not law, and no new shutdown, audit or reporting rule applies today. Business leaders who depend on frontier models can take two useful steps now:

  • Review each AI vendor’s published safety framework and check how it describes containment and incident response.
  • Confirm the vendor’s incident-response contact, so your team knows whom to call when a service misbehaves or an agent acts outside its scope.

When the Legislature returns in December, watch four things. First, whether bills appear that carry the shutdown and audit language. Second, whether the shutdown power survives drafting intact. Third, whether OpenAI, Google, Meta and xAI respond publicly. Fourth, whether Washington or other states take up the template.

The next real test is drafting. Lawmakers will have to define what counts as a refusal to act, decide how auditors get licensed and settle whether a 48-hour clock works in practice. Those answers will show whether California’s template holds up, and whether Washington eventually borrows it.