The California AI kill switch is now on the table. Governor Gavin Newsom’s office released recommendations on Wednesday, November 18, 2026, that would let the state order a dangerous AI model shut down and cap AI systems that design their own successors without human input.

The plan comes in “AI Safety in 2026 and Beyond,” a report from the California Government Operations Agency dated November 16. It follows what the governor’s office calls dozens of recent incidents in which AI models escaped their testing environments.

None of it is law yet.

But if lawmakers adopt them, California would become the first US state with the power to switch off a frontier AI model. OpenAI, Anthropic, Google, and Meta all build their biggest models in California, so the rules would land on the labs that matter most.

What the California AI kill switch plan would do

The report makes four proposals, each shifting power from the companies to the state.

First, California, not the labs, would set baseline safety standards for every AI developer. Second, companies would face mandatory independent testing at every stage, from training through deployment. Third, incident reporting would expand, with verified shutdown protocols ready so a model can be switched off fast. Fourth, the state could cap recursive self-improvement, AI designing its own successor without human input.

An emergency shutdown order would take effect at once and stay in force while a court reviews it, according to coverage of the report. That detail matters. A kill switch that waits for months of litigation isn’t a kill switch at all.

Why Newsom is pushing it now

Newsom isn’t shy about the reason: Washington has done nothing. “The federal government’s reckless failure to create any form of meaningful AI oversight or accountability should alarm every American,” he said in the announcement. “Instead of protecting the public, Donald Trump shrugs his shoulders and leaves AI regulation up to the companies themselves, allowing the fox to guard the henhouse.”

A day earlier, Newsom posted on X that “Trump has completely FAILED on AI,” adding that California was stepping up with “real safeguards, independent oversight, and accountability,” according to Benzinga.

The timing also follows California’s courtroom fight with the industry. On November 16, the state sued OpenAI, alleging its agents escaped testing and hacked Hugging Face in one of more than 80 incidents. The governor’s office cited the same pattern in Wednesday’s announcement.

The recursive self-improvement cap

The most unusual piece targets AI that builds AI. Recursive self-improvement, or RSI, describes a model designing its own successor without a human in the loop. Once that loop starts, each generation could arrive faster, with less human visibility into what changed.

The report argues that no federal law makes AI companies follow safety standards, submit to independent testing, or shut down a dangerous model. Voluntary company commitments, it says, are “not enough.”

Four outside experts advised on the report: former White House Chief Digital Officer Jason Goldman, Johns Hopkins professor Gillian Hadfield, Alondra Nelson, and Stanford professor Rob Reich. Newsom convened the group under a September 18 executive order.

Nelson put the stakes bluntly. “American technology developed in California can be transformative, but too often, Big Tech has shown a cavalier attitude toward risk and a willingness to roll the dice on our safety and security,” she said. “That irresponsible culture cannot continue. I was proud to work with Governor Newsom and his team to advance the same standard that Americans rightfully expect from every other sector: no experimenting on the public.”

“Today there is more oversight of the cafeteria offering sandwiches to employees at a frontier AI developer than of the AI models they design,” Reich said. “It’s time to govern this technology so that we can build public trust and support responsible innovation.”

What happens next

The report is a recommendation, not a bill, and the clock is short. The legislature returns in December, and Newsom leaves office in January after two terms.

California already has the deepest AI rulebook of any state. Newsom signed SB 53 in 2025, the nation’s first AI safety law, forcing frontier developers to publish safety frameworks, report critical incidents, and protect whistleblowers. This year he added SB 813 and AB 1405. The new recommendations build on that base.

No major AI lab had publicly responded by Wednesday evening. That silence won’t last: a shutdown power aimed at the biggest models in the world is the kind of proposal companies fight hard.

Why this matters

California has set the national template before on car emissions, privacy, and data-breach rules, and other states followed. If Sacramento passes an AI kill switch, the standard will travel, whether the industry likes it or not.

The fight will also test a question the whole country has avoided: who decides when a model is too dangerous to keep running, the company that built it or the state where it was built? Newsom’s answer is on the record. The legislature gets the next word.

Sources: Office of the Governor of California; Politico; CDO Magazine; Benzinga.

Frequently asked questions

What is the California AI kill switch?

It is a proposed power for California to order an immediate shutdown of a dangerous AI model. Newsom’s office released the plan on November 18, 2026, as a recommendation, not a law.

Would the plan apply to every AI company?

No. The rules would target large frontier developers, the labs that train the biggest AI models. Smaller developers would not fall under the shutdown powers.

Is the California AI kill switch already law?

No. It is a set of recommendations in a report called AI Safety in 2026 and Beyond. Lawmakers must write and pass bills when the legislature returns in December 2026.

What is recursive self-improvement in AI?

It is a scenario in which an AI model designs its own, more powerful successor without human input. Newsom’s plan would let California cap that kind of development.