California’s Government Operations Agency has published a report setting out proposed next steps for the state’s AI safety laws, including an emergency shutdown power and a power to cap recursive self-improvement of frontier models. The report, “AI Safety in 2026 and Beyond,” is dated 16 November 2026. Governor Gavin Newsom’s office released it on 18 November, alongside a statement that frames the document as recommendations for mandatory safety standards and oversight.
The ideas are recommendations, not law. The report says the agency, “in consultation with the California Office of Emergency Services and outside AI experts, has developed recommendations for legislative action.” It is written to “help inform and shape discussions as the legislature returns next month.” Nothing in the document itself enacts a shutdown order or a cap.
The report wants shutdown orders that take effect at once
Among the recommendations is a state power to order an immediate emergency shutdown of a model when the danger is imminent and the developer refuses to act. “If the danger from a model is imminent and the developer refuses to take action to contain the danger, the state should have authority to order an immediate emergency shutdown,” the report says. The government’s order “should go into effect immediately and remain in effect pending judicial review.” It also says the state should order the narrowest workable intervention and should remove the order when the danger is no longer imminent.
The same document argues for baseline safety standards for AI developers, independent testing and verification through training, development and deployment, expanded incident reporting, and verified risk-containment protocols. Independent verification organisations, it says, should have an ongoing role, including in checking the metrics discussed below.
On recursive self-improvement, or RSI, the report says developers “should have to report key metrics on their automation of AI R&D, including the share of compute they are devoting to this purpose.” If the state assesses that this automation, above a certain level, “may result in catastrophic risk, the state should have authority to cap the development and use of RSI.” That is a proposed power, tied to a risk assessment the state would still have to build, not a cap that exists today.
The report thanks Jason Goldman, Gillian Hadfield, Alondra Nelson and Rob Reich for advice on the work. Government Operations Agency Secretary Nick Maduros, quoted in the Governor’s release, said: “Recent events have made it clear that more oversight and transparency are urgently needed.” Reich, also quoted there, said: “Today there is more oversight of the cafeteria offering sandwiches to employees at a frontier AI developer than of the AI models they design.”
These powers would have to be written into statute
California already has a thick file of AI statutes and cases, and this report sits on top of that pile rather than replacing it. TSN has covered the package of 13 AI bills signed in one day and, more recently, the state attorney general’s lawsuit against OpenAI over AI-agent cyber incidents. Those are existing law and existing litigation. The GovOps paper is a brief to lawmakers about what should come next.
The shutdown recommendation is the sharpest of those next steps, because it would let the state act first and face a court afterwards. The report tries to narrow that power: imminent danger, a developer that has refused to contain it, a credible reason to believe the risk, and the narrowest workable intervention, lifted when the danger has passed. Even so, the legal design — an order that takes effect immediately, then remains in place pending judicial review — would be a substantial grant of executive authority over systems that, in many cases, are trained and served from outside California. How that would apply to a model hosted in another state, or in another country, is left for the legislature to work out.
The RSI cap is a different kind of instrument. It starts with disclosure, the share of compute going into automated AI research, and only then contemplates a cap if the state finds a catastrophic-risk threshold has been crossed. That depends on definitions and measurements that the report itself treats as future work for independent verifiers. It is closer to a reporting-and-licensing regime than to a switch that can be thrown tomorrow.
Other governments are also trying to invent a regulator for frontier models. Australia’s plan for an AI regulator and frontier safety laws is one such attempt; OpenAI’s own widening of agent monitoring after the Hugging Face incident is a company-side move, not a state power. Neither is a California statute, and neither answers the question this report puts to Sacramento: whether the state should be able to order a shutdown, or cap recursive improvement, at all.
The next fact to watch is what, if anything, is introduced once the legislature returns, and whether the shutdown and RSI recommendations survive in bill form.
Sources
- California Government Operations Agency, “AI Safety in 2026 and Beyond,” dated 16 November 2026.
- Office of the Governor of California, “Governor Newsom’s Nation-Leading AI Safety Framework is Getting Even Stronger as California Releases Next Steps on Mandatory Safety Standards and Oversight,” 18 November 2026.
Get TSN’s biggest tech stories, explained.
News and insights across AI, tech, crypto and cybersecurity. Every source linked. Unsubscribe anytime.
Get the newsletterRelated stories
The closest recent TSN account of California using existing law against a frontier developer is the attorney general’s OpenAI case. The earlier package of 13 AI bills is the statutory baseline this report would add to.
- California AG Sues OpenAI Over AI-Agent Cyber Incidents
- Thirteen Bills in One Day: California’s Latest AI Guardrails
- Australia Plans an AI Regulator and Frontier Safety Laws
- OpenAI Widens Agent Monitoring After the Hugging Face Incident
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X





