Regulation

Newsom Releases AI Kill Switch Plan With 48-Hour Rule

California's AI safety report cuts incident reporting from 15 days to 48 hours and proposes emergency shutdown powers for frontier models.

Share:XLinkedIn

Key Takeaways

  • 48-hour reporting window proposed: California wants to cut critical incident reporting from 15 days to 48 hours, aligning AI emergencies with the state's existing data breach disclosure standards.
  • Kill switch requires new legislation: The November 18 report is a recommendations document, not a law; Governor-elect Becerra must introduce and pass bills before any shutdown authority becomes enforceable.
  • SB 813 and AB 1405 already signed: Independent verification organizations and an AI auditor registry were enacted September 9, 2026, with operational deadlines in mid-to-late 2027.
  • xAI challenged a less aggressive California law the same day: The Ninth Circuit was skeptical of xAI's bid to block AB 2013, which only requires training data summaries; mandatory kill switch authority will face stronger legal challenges.
  • No federal parallel exists: The Trump administration's AI framework is voluntary; California is the only U.S. jurisdiction currently proposing binding emergency shutdown authority for frontier AI systems.

California Governor Gavin Newsom released a formal policy report on November 18 calling for state officials to gain emergency shutdown authority over frontier AI systems, with a companion proposal to cut the critical incident reporting window from 15 days to 48 hours. The report, produced by a four-person expert panel commissioned through an executive order Newsom signed in September, represents the most detailed government blueprint for regulating frontier AI models published by any U.S. state. It arrives in the same week that California's attorney general filed a lawsuit against OpenAI, the Ninth Circuit heard arguments over the state's AI training data law, and the federal government signaled it has no plans for binding AI oversight of its own.

What the Report Actually Proposes and What Is Already Law

The "AI Safety in 2026 and Beyond" report, prepared by the California Government Operations Agency and the Office of Emergency Services, is not a bill and does not create new legal obligations on its own. It is a set of recommendations that the state government says it intends to pursue through legislation. Governor-elect Xavier Becerra, who takes office in January, has pledged to move quickly to put "full incident reporting and real shutdown authority into law," according to the governor's official press release. Legislators return to Sacramento on December 7, giving the new administration roughly two months before it can begin introducing legislation.

The underlying framework that makes these proposals possible was set in motion by Executive Order N-9-26, which Newsom issued on September 18. That order gave the Government Operations Agency a November 16 deadline to deliver recommendations on four specific questions: whether to place independent auditors onsite at AI labs, whether to verify the safety frameworks companies already file under state law, whether a kill switch is technically feasible, and how to expand the category of incidents companies must report. The November 18 report is the formal answer to those questions, arriving two days after the original deadline. Some of the structural scaffolding the report assumes is already in place. Senate Bill 813, which Newsom signed on September 9, 2026, directs the Government Operations Agency to designate independent verification organizations as outside auditors qualified to assess AI safety risks. The Assembly passed it 67 to 6, and the Senate concurred 39 to 0, making it one of the least contested major tech bills in recent California legislative history. Assembly Bill 1405, also signed September 9, creates a state registry for those auditors with standards for independence and integrity, as reported by Bloomberg Government. Under the governor's September signing announcement, state agencies must post qualification criteria for verification organizations by May 1, 2027. The auditor registry must be operational by December 1, 2027.

From 15 Days to 48 Hours: The Incident Reporting Change

The proposal to compress the reporting window from 15 days to 48 hours may ultimately matter more in practice than the kill switch headline. Under Senate Bill 53, California's existing AI transparency law, frontier developers must report critical safety incidents to state officials. SB 53 sets that window at 15 days. The new report recommends cutting it to 48 hours. The same 48-hour standard applies in California to the most serious data breaches under existing privacy law, so the proposal aligns AI incident reporting with the state's established emergency-response expectations rather than inventing a novel framework. The practical effect is that an AI lab discovering a loss-of-control incident at noon Monday would have until noon Wednesday to disclose it to regulators, rather than three weeks.

Stay Ahead

Get daily AI signals before the market moves.

Join founders, investors, and operators reading TechFastForward.

The report also recommends expanding which incidents trigger that reporting obligation. Under the current definition in state law, the threshold for a reportable event is catastrophic harm: a single incident capable of killing more than 50 people or causing $1 billion in theft or property damage through a cyberattack or an autonomous system acting outside its intended scope. The panel wants to add loss-of-control incidents as a separate reportable category: cases where a model behaves in ways the company cannot stop or reliably predict, even if no catastrophic harm has yet occurred. That change is a broad expansion: it would force disclosure of model misbehaviors before they reach the existing threshold, giving regulators an earlier signal rather than a post-incident audit trail.

Whistleblower protections currently under SB 53 cover employees of frontier AI developers. The report recommends extending them to contractors and former employees, closing a gap that critics have flagged since SB 53 was enacted. The funding mechanism for the broader oversight system is also proposed here: a surcharge on the fees that AI companies pay to state-licensed auditors would fund CalCompute, the state's public cloud initiative. No dollar amount is specified in the report, but the design is intentional: it ties the cost of oversight directly to the regulated industry rather than the state general fund.

Newsom Is Filling a Federal Vacuum

The timing and framing of the report cannot be separated from the federal backdrop. The Trump administration has taken a voluntary approach to AI oversight: a White House framework announced in mid-2026 invites AI companies to submit frontier models for voluntary government review before release, but it carries no binding requirements and has not been released publicly. No federal legislation mandating safety reporting, independent auditing, or emergency shutdown authority for AI systems has advanced through Congress. Newsom framed the gap in explicit terms: "Instead of protecting the public, Donald Trump shrugs his shoulders and leaves AI regulation up to the companies themselves," he said at the report release, as reported by CalMatters.

California's posture now looks more like a parallel federal regime than a supplement to one. The state already requires frontier AI developers to publicly disclose their safety frameworks, report critical incidents, and protect whistleblowers under SB 53. It has established a certification pathway for independent auditors through SB 813 and AB 1405. The November 18 report adds proposed layers: emergency shutdown authority, faster incident reporting, state-assigned evaluators, and limits on recursive self-improvement, the term the report uses for AI systems designing their own successors without human input. No other U.S. state has proposed regulatory controls on that specific development pathway, and no federal proposal in the current Congress has either. The EU AI Act, which entered the enforcement phase in August 2026 with fines of up to 3 percent of global annual revenue or 15 million euros, whichever is higher, also does not treat autonomous AI self-improvement as a regulated category.

The expert panel behind the recommendations includes Alondra Nelson, former acting director of the White House Office of Science and Technology Policy and now a faculty member at the Institute for Advanced Study; Rob Reich of Stanford University; Jason Goldman, former White House Chief Digital Officer and head of the Center for Shared AI Prosperity; and Gillian Hadfield of Johns Hopkins University. The panel's composition spans institutions and party affiliations, which Newsom's office has used to insulate the recommendations from criticism that they represent a narrow ideological preference. Rob Reich's framing at the report launch captured the core argument: there is more state oversight of a frontier AI developer's employee cafeteria than of the models it designs and releases to hundreds of millions of users.

What the Ninth Circuit Fight Tells You About Lab Resistance

On the same day the report was published, a Ninth Circuit panel heard arguments from Elon Musk's xAI challenging California's Assembly Bill 2013, which requires AI developers to publish summaries of the data they used to train their models. xAI argued the law compels speech in violation of the First Amendment. The appeals panel was openly skeptical of that position, with Judge Johnnie Rawlinson disputing xAI's contention that the disclosure requirement amounts to unconstitutional compelled commercial speech. A federal district court had already denied xAI's request for a preliminary injunction in March, leaving AB 2013 in force while the appeal runs. No ruling has been issued from the Ninth Circuit yet.

The xAI case is a preview of the legal environment surrounding far more aggressive measures. AB 2013 requires only a summary of training data, the least intrusive form of AI disclosure California has enacted. The November 18 report recommends requirements that go considerably further: mandatory access for state-designated auditors to lab facilities, verified shutdown capabilities that auditors independently test, and reporting of incidents that do not yet qualify as catastrophic harms under existing law. Any legislation enacting those requirements will almost certainly face the same First Amendment and federal preemption arguments that have followed AB 2013. The Department of Justice joined xAI's challenge to Colorado's AI Act earlier in 2026, signaling federal willingness to push back on state AI laws the administration views as overreaching, which is an irony given that administration's simultaneous opposition to federal AI regulation. That precedent will follow every bill Becerra introduces in Sacramento.

The Kill Switch Cannot Work the Way the Report Describes It

The bear case for the kill switch is not philosophical: it is technical and jurisdictional. California can require that companies operating within its borders maintain and demonstrate the ability to shut down a frontier model in an emergency. What it cannot do is guarantee that the same model, deployed in cloud infrastructure in Virginia or served through a data center in Singapore, actually goes dark when Sacramento says so. Cloud-native AI systems are distributed by design, and the most commercially deployed models from OpenAI and Anthropic run across dozens of geographic regions simultaneously. A California emergency order directed at a lab headquartered in San Francisco carries no automatic authority over compute located outside the state, let alone outside the country.

Critics also point to the auditor independence problem the report itself acknowledges. Rif Saurous, a staffer at METR, the AI safety evaluation organization, told a state Senate hearing that independent auditors are "under soft pressure to maintain good relationships with the labs." The report's answer is to let the state assign evaluators to labs rather than allowing AI companies to hire their own auditors, but the pipeline to get there is long. The Government Operations Agency must post qualification criteria by May 1, 2027, begin accepting applications, certify the first cohort, and then assign them before any mandatory onsite evaluations can begin. That timeline puts the first enforcement activity in mid-to-late 2027. Anthropic's chief global affairs officer Tino Cuellar offered a carefully worded response: the company is "encouraged by the report's push on independent evaluation," a formulation that stops short of committing to voluntary compliance before legislation requires it. OpenAI did not issue a public response to the report.

Becerra's Pledge and What to Watch Before March

Governor-elect Becerra's pledge to introduce legislation on incident reporting and shutdown authority gives the November 18 report its first legislative test. Legislators return to Sacramento on December 7, 2026. Bills of this profile typically receive January committee hearings, putting the first real test of Becerra's commitment in the first weeks of the new year. Two specific signals matter. First, whether the incident reporting change, compressing the window from 15 days to 48 hours, is introduced as a standalone amendment to SB 53 or packaged with the broader emergency shutdown authority. The two have very different legal and political trajectories: the reporting change has clear precedent in existing California privacy law and is unlikely to draw industry opposition, while the kill switch authority will almost certainly trigger First Amendment challenges from labs that prefer to manage their own safety disclosures.

Second, watch whether the limits on recursive self-improvement make it into the first bill or are deferred to a later session. No existing AI law anywhere in the world treats autonomous AI self-improvement as a regulable category. California would be writing on a blank slate, which carries both regulatory risk and strategic leverage: a California rule would become the de facto global template in the absence of federal or international alternatives, in the same way that the state's vehicle emissions standards have historically set the floor for automotive regulation across much of the country. State Sen. Scott Wiener, the architect of several California AI safety bills including SB 53, is heading to Congress, where he has pledged to pursue federal AI safety legislation after he is sworn in. If his federal efforts gain traction, the pressure on California's most aggressive proposals will ease and the state's rules can serve as a floor rather than a ceiling. If those federal efforts stall, California will be left as the sole regulatory backstop for AI systems that are global in scope but legally anchored to a single state.

California cannot guarantee a kill switch shuts down a model that lives in a data center it doesn't govern, but the law doesn't have to be perfect to reshape how labs build for safety from day one.


Key Takeaways

  • 48-hour reporting window proposed: California wants to cut critical incident reporting from 15 days to 48 hours, aligning AI emergencies with the state's existing data breach disclosure standards.
  • Kill switch requires new legislation: The November 18 report is a recommendations document, not a law; Governor-elect Becerra must introduce and pass bills before any shutdown authority becomes enforceable.
  • SB 813 and AB 1405 already signed: Independent verification organizations and an AI auditor registry were enacted September 9, 2026, with operational deadlines in mid-to-late 2027.
  • xAI challenged a less aggressive California law the same day: The Ninth Circuit was skeptical of xAI's bid to block AB 2013, which only requires training data summaries; mandatory kill switch authority will face stronger legal challenges.
  • No federal parallel exists: The Trump administration's AI framework is voluntary; California is the only U.S. jurisdiction currently proposing binding emergency shutdown authority for frontier AI systems.

Questions Worth Asking

  1. If OpenAI or Anthropic refuse a California emergency shutdown order, what enforcement mechanism does the state actually have against a system running on cloud infrastructure in multiple countries?
  2. Would a mandatory 48-hour incident reporting window create incentives for AI companies to suppress internal safety findings that do not yet meet the legal threshold for a required report?
  3. If California's limits on recursive self-improvement set a global template, does that make the state the de facto regulator of a technology that no jurisdiction yet knows how to measure?

Sources

  1. official press release gov.ca.gov
  2. Bloomberg Government news.bgov.com
  3. September signing announcement gov.ca.gov
  4. CalMatters paloaltoonline.com
  5. openly skeptical of that position mlq.ai

Read Next

SpaceX Sells $32B in Bonds to Finance Nvidia Rubin Chips

Anthropic Raises $119.5B in Record IPO, Stock Soars 17%

Meta Plans $300 Muse Charm With Free Verizon 5G Data

DJI Air 3S Outranges Neo by 25 km and Costs 5.5x More

Newsletter

Enjoyed this analysis? Join the list for the email edition.

Daily AI signals. No noise. Built for founders, investors, and operators.

Share:XLinkedIn
</> Embed this article

Copy the iframe code below to embed on your site:

<iframe src="https://techfastforward.com/embed/newsom-releases-ai-kill-switch-plan-with-48-hour-rule" width="480" height="260" frameborder="0" style="border-radius:16px;max-width:100%;" loading="lazy"></iframe>