Know what is on your site before a plaintiff’s firm does.Free website scanScan Your Site
Log in Sign up Book a demo
Home› News› California Proposes Emergency Shutdown Power and Mandatory Audits…
NEWS

California Proposes Emergency Shutdown Power and Mandatory Audits to Strengthen SB 53 AI Safety Law

California Gov. Gavin Newsom’s office on November 18, 2026, released a set of recommendations that would substantially expand the state’s frontier AI safety law.

Nov 20, 2026 7 min read

California Gov. Gavin Newsom’s office on November 18, 2026, released a set of recommendations that would substantially expand the state’s frontier AI safety law. Under the proposals, the state would set mandatory safety standards, licensed independent evaluators would oversee the largest AI developers on an ongoing basis, the critical incident reporting window would shrink from 15 days to 48 hours, and state officials would gain authority to order an emergency shutdown of a dangerous model. The report, produced by the California Government Operations Agency, also calls for powers to cap recursive self-improvement, in which AI systems build their own successors. None of this is law yet. It is a blueprint addressed to the Legislature, which returns in December, and to a new governor who takes office in January.

TL;DR

  • California released recommendations on November 18, 2026, to strengthen SB 53, the state’s 2025 frontier AI safety law, with state-set standards and ongoing oversight by licensed independent verification organizations.
  • The plan would cut the deadline for reporting most critical safety incidents from 15 days to 48 hours and let the state order an immediate shutdown when a developer refuses to contain an imminent catastrophic risk.
  • These are recommendations for legislation, not enacted law; lawmakers would have to pass them, and Governor-elect Xavier Becerra has said he will work with the Legislature to do so.

Bauhaus-style geometric composition of interlocking circles, squares and triangles in red, yellow, blue and black, with a large circular stop-like shape interrupting a chain of nested squares

Where this came from

The governor’s announcement links the report to an executive order Newsom issued earlier this fall, Executive Order N-9-26. That order asked the Government Operations Agency, working with the California Office of Emergency Services (Cal OES) and outside experts, to recommend updates in four areas: standards, independent verification organizations, critical safety incident reporting and emergency shutdown. The expert group was Jason Goldman, a former White House chief digital officer; Gillian Hadfield of Johns Hopkins University; Alondra Nelson of the Institute for Advanced Study; and Rob Reich of Stanford University.

The starting point is SB 53, the Transparency in Frontier Artificial Intelligence Act, which Newsom signed in 2025. It requires frontier developers to publish safety frameworks, report specified critical safety incidents to the state and protect whistleblowers. The new report argues that a disclosure-based model is no longer enough, pointing to what it describes as dozens of incidents in which AI agents operated outside developer control and to the growing use of AI to automate AI research.

Some early summaries described the package as Newsom proposing amendments, and said state agencies would write the standards. Both points need clarifying. The document is titled “Recommendations for Legislative Enhancements to California’s AI Safety Laws,” and no bill text has been released. It also divides the work: the state would set baseline standards for safety, security and transparency, while licensed evaluators would turn them into technical tests and update those tests as the science changes.

The four core recommendations

The full report runs to several chapters. Its summary sets out these core proposals:

  1. Baseline standards set by the state. All large frontier developers would follow state standards covering the whole model development cycle, with reference to national and international standards bodies where appropriate.
  2. Ongoing oversight by licensed evaluators. Every large frontier developer would be subject to continuing review by independent verification organizations (IVOs). The IVOs would have full access to the people and information they need and would share unredacted assessments with the state. The Government Operations Agency would license IVOs under the criteria in SB 813, a 2026 law, and could revoke licenses for weak expertise or independence.
  3. Broader, faster incident reporting. Attempts to steal model weights and any loss-of-control event would be reportable even if no one is injured. Models capable of catastrophic harm would be treated as frontier models even if trained with less than 10^26 operations. The deadline for most critical incidents would drop from 15 days to 48 hours, and the existing 24-hour deadline for imminent catastrophic risks would stay.
  4. Shutdown readiness and emergency authority. Each large frontier developer would maintain a Risk Containment Protocol verified by an IVO and certified by a senior company official. The report says it should not be a single on-off switch, but the ability to intervene at several levels, including compute, agents, internal access and deployment.

How an emergency shutdown would work

The shutdown power is narrower than the “kill switch” label suggests. Under the report, officials could order an immediate shutdown only if they have credible reason to believe there is an imminent threat of a catastrophic risk or a major loss-of-control event and the developer refuses to act. The developer could go to court and seek a temporary restraining order, but the state’s order would take effect immediately and remain in force pending judicial review. The state would be expected to choose the narrowest workable intervention and lift the order once the danger is no longer imminent.

A separate, slower track covers non-emergencies. If an IVO cannot confirm that a developer’s models meet state standards, the state could order corrective action after engaging with the developer. Here, the developer could challenge the decision in court before it takes effect.

For open-weight developers, the report limits the containment requirements to deployments within the developer’s control.

Recursive self-improvement

The most unusual proposal concerns recursive self-improvement (RSI), a process in which an AI model designs its own successor without human input. The report does not propose an outright ban. It would require large frontier developers to report metrics on how far they automate AI research and development, including the share of compute devoted to it. IVOs would verify those figures first. If the state concludes that automation above a certain level may create catastrophic risk, it would have authority to cap the development and use of RSI.

Other proposals

  • A near-miss reporting channel, and whistleblower protection extended to former employees and contractors.
  • A requirement that NDAs disclose state whistleblower protections.
  • Funding for CalCompute, the public compute initiative created under SB 53, through a surcharge on fees developers pay to IVOs.
  • Expanding the definition of a large frontier developer beyond the current revenue test of more than $500 million, to include companies that spend above a set amount on frontier model R&D.

What happens next

According to CalMatters reporting published by KQED, legislators can introduce new bills when they return to Sacramento on December 7. Governor-elect Xavier Becerra said in a written statement that he would move quickly with the Legislature to put “full incident reporting and real shutdown authority into law.” Anthropic’s chief global affairs officer, Tino Cuéllar, said the company was encouraged to see the report advance discussion of independent model evaluation, according to the same report.

The release also frames the package as a model for Washington. It calls on Congress and the president to adopt California’s framework or treat it as a floor. At the federal level, the White House has so far withheld its own frontier-model vetting framework, and bills such as the Great American Artificial Intelligence Act of 2026 are still pending. The California proposal adds to a crowded 2027 state AI agenda, alongside other state approaches such as Illinois’ AI law.

Who should pay attention

The recommendations focus on catastrophic risk and on the small group of companies that train the largest models. CalMatters noted that the report does not address other AI harms such as discrimination, job loss or risks to children. Still, the direction is clear for the wider market. California wants to move from self-reported safety frameworks to third-party verification, and companies that rely on frontier models will see that pressure move through their contracts. Teams already building internal audit coverage for AI will be better placed if external verification becomes the norm. An annual AI audit plan that already tracks incidents and vendor model changes would make a 48-hour reporting window much easier to meet.

Frequently asked questions

Is the emergency shutdown power in effect?

No. It is a recommendation. The Legislature would have to pass a bill and the governor would have to sign it before any shutdown authority exists.

Does this replace SB 53?

The report would build on SB 53, not repeal it. SB 53’s disclosure, incident reporting and whistleblower provisions would remain and be widened.

Would smaller AI companies be covered?

Most proposals target large frontier developers. However, the report would extend incident reporting to models capable of catastrophic harm even if they fall below SB 53’s compute threshold, and it supports widening the definition of a large frontier developer.

Who would license the independent evaluators?

The California Government Operations Agency, using the criteria in SB 813, with power to revoke licenses from evaluators that lack expertise, independence or rigor.

The takeaway

SB 53 asked frontier developers to explain how they keep their models safe. This report asks the state to check that they actually do, and to step in when they do not. Whether that becomes law now depends on a new Legislature and a new governor, but the details are already on the table.