California Proposes 48-Hour AI Incident Reporting and State Kill Switch

Newsom's November 18 plan would cut AI incident reporting from 15 days to 48 hours, license state auditors, and let officials shut down frontier models like OpenAI's and Anthropic's.

By Grace Kim

4 min read

California Proposes 48-Hour AI Incident Reporting and State Shutdown Authority, But Nothing Is Law Yet
California Proposes 48-Hour AI Incident Reporting and State Shutdown Authority, But Nothing Is Law Yet12019 / pixabay

What's News

  • California's November 18 report proposes cutting critical AI incident reporting from 15 days to 48 hours.
  • State officials could order immediate shutdowns of frontier models from companies like OpenAI and Anthropic.
  • The 85-page report responds to Newsom's September 18 executive order demanding an 'AI kill switch' within two months.
  • Lawmakers return December 7; the recommendations are proposals, not enacted law.
  • Catastrophic risk is defined as an event killing 50+ people or causing $1 billion in damage.

Governor Gavin Newsom on November 18 released an 85-page legislative blueprint that would cut California's critical AI safety incident reporting deadline from 15 days to 48 hours and give state officials power to order emergency shutdowns of frontier AI models. None of it is law yet. The recommendations are meant to inform bills when the Legislature returns to Sacramento on December 7.

The report, titled AI Safety in 2026 and Beyond, came from the California Government Operations Agency working with the Office of Emergency Services and four outside advisers: Jason Goldman, Gillian Hadfield, Alondra Nelson, and Rob Reich. It answers a September 18 executive order in which Newsom demanded proposals for independent oversight and an "AI kill switch" within two months.

What would the shutdown authority actually do?

State officials could order an immediate shutdown when they have "credible reason to believe there is an imminent catastrophic risk or major loss-of-control event and the developer refuses to contain the danger." The order must be the "narrowest workable intervention," takes effect immediately, and stays in force during judicial review. Developers could still seek a temporary restraining order. The power would apply to companies such as OpenAI and Anthropic.

How fast would companies have to report incidents?

  • Most critical safety incident reports: 15 days cut to 48 hours
  • Imminent catastrophic risks: existing 24-hour requirement unchanged
  • New reporting scope: attempted model-weight theft, all internal or external loss-of-control incidents, and a new near-miss channel
  • Whistleblower protections extended to contractors and former employees

The takeaway: the state wants to know about near-disasters within two days, not two weeks.

Who watches the developers?

Every large frontier developer would face continuous monitoring by state-licensed Independent Verification Organizations (IVOs). The Government Operations Agency would license these auditors and could revoke licenses for insufficient expertise, independence, or rigor. Crucially, the state—not the AI companies—assigns evaluations to auditors. Auditors would get access to models, personnel, systems, training environments, and records, and would send unredacted assessments to the state.

Each large frontier developer would also maintain a Risk Containment Protocol covering compute, agents, internal access, and deployment—verified by an IVO and certified by a senior company official. Testing would include failure cases such as a model copying its own weights to resist containment. Open-weight developers would be responsible only for deployments they control.

What about self-improving AI?

Officials could impose a time-limited, periodically reviewed cap on automated AI research when it exceeds a developer's committed thresholds or advances faster than its safeguards and evaluations, creating a material risk of catastrophic harm. The cap would lift once an IVO verified adequate safeguards. Separately, companies automating AI R&D would report quarterly how their compute splits across training, research experiments, inference for AI-research agents, safety and security work, external customers, and unattributed activity, with IVOs verifying the categories.

The report also redefines what counts as "frontier." Models capable of actions that could create catastrophic risk would fall under reporting rules even if trained with fewer than 10^26 integer or floating-point operations—a direct answer to the concern that training-scale thresholds alone have left dangerous incidents outside state reporting requirements.

A surcharge on the verification fees developers pay to auditors would help finance CalCompute, the state's public cloud initiative, with some capacity reserved for independent safety research and public oversight.

What's the political path from here?

Newsom framed the effort as filling a vacuum left by federal inaction. "Instead of protecting the public, Donald Trump shrugs his shoulders and leaves AI regulation up to the companies themselves," he said in a press release. Governor-elect Xavier Becerra said in a written statement that he will move quickly with the Legislature to put "full incident reporting and real shutdown authority into law."

State Senator Scott Wiener, who wrote a prior state law on catastrophic AI dangers and was recently elected to Congress, has pledged to introduce AI safety regulation after he is sworn in. The recommendations build on existing law: SB 53 (2025), the Transparency in Frontier Artificial Intelligence Act, requiring disclosure of safety frameworks, incident reporting, and whistleblower protections; plus SB 813 and AB 1405 (2026), which set up certification for independent verification organizations and a state registry for AI auditors.

Anthropic's chief global affairs officer, Tino Cuéllar, said the company is "encouraged to see the report move the conversation forward on issues like the independent evaluation of AI models" and will "continue working with leaders in California and throughout the world to make sure we keep making progress on these issues at a critical time." No reaction from OpenAI was reported.

What the report leaves out

The report covers only catastrophic risk—defined in state law as a single event capable of killing more than 50 people or causing $1 billion in theft or property damage through a cyber attack or autonomous system activity. It skips automated discrimination, job loss, and threats to children's safety or mental health. The coauthors argue that building a credentialed-auditor system for catastrophic risk will make it easier to evaluate models for other harms later.

These are expert proposals, not binding requirements. Whether the Legislature enacts any, all, or none of them—and on what timeline—remains the open question heading into December 7.

Source: epyhia.com

Share this article:

More from Grace Kim

Grace Kim

Show full bio

Market editor covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.

2,419 articles

Related articles

« Previous articleNext article »