California Governor Gavin Newsom has roughly two months to turn a September executive order into recommendations that could lead to a state “kill switch” requirement for powerful AI models. The Newsom AI kill switch idea is back in the news because Al Jazeera published a profile on October 17 that ties his AI push to a possible 2028 presidential run.

The order itself is not new. Newsom signed it on September 18, a Friday, and it directs state officials to develop recommendations for AI safety legislation. What has changed is the calendar and the attention. The recommendations are due in mid-November, and any mandatory shutoff rule would reach AI companies whose models are used across the country, not only inside California. No new state action surfaced on October 16 or 17 beyond that coverage, so this is a follow-up rather than a fresh announcement.

What the order requires, and what it only proposes

The order directs the California Government Operations Agency, working with national experts, to develop recommendations for AI safety legislation, according to Daily Bruin and StateScoop. Most of the ideas on the table are still proposals. StateScoop reports that the options under consideration include:

  • Requiring large frontier developers to have independent verification organizations audit them periodically.
  • An emergency “kill switch” that could shut down a model, tested regularly by independent groups.
  • Broader definitions of “reportable critical safety incidents” to cover AI systems that lose control of their operations.

The order also covers two recently signed California laws. One is SB 813, which concerns independent AI safety evaluation groups. The other is AB 1405, which sets auditor standards and creates a registry, according to Quartz.

Quartz also reports that the order does not immediately require AI companies to build a shutoff capability. That distinction matters. A recommendation is not a rule. Acting on these ideas would require the Legislature, which Newsom could convene in a special session, possibly before his term ends, according to Al Jazeera.

When the recommendations are due, and who is doing the work

Three outlets, Daily Bruin, StateScoop and Quartz, put the recommendation deadline at November 16. Al Jazeera reports that Newsom asked the California Emergency Services Office to send recommendations on AI regulation by November 15. The one-day gap is unresolved in the reporting, so anyone planning around the date should check the order text before relying on either one.

The expert work has its own clock. NBC News reports that a plan is due within two months. Quartz says the panel is examining independent monitors inside AI labs, company-level safety planning requirements, and an automatic shutdown triggered by critical safety failures. StateScoop reports that application requirements and procedures for independent verification organizations are due by next May, with implementation by the end of 2027.

Rebecca Finlay, chief executive officer of Partnership for AI, a nonprofit focused on technology governance, described the moment this way in Al Jazeera’s reporting: “We are building the field of AI safety in real time.”

How the Newsom AI kill switch idea differs from the SB 1047 veto

Newsom has taken this route before, and he stopped it. In 2024 he vetoed SB 1047, a bill that would have required developers to be able to fully shut down advanced AI systems on demand. He cited innovation concerns, according to Quartz. The new order revives that concept, but it is narrower in one respect. It covers shutdown mechanisms, while the vetoed bill required safeguards before training began, as Al Jazeera notes.

That shift matters for who bears the cost. A training-stage rule touches a lab’s development process. A shutdown rule touches systems already running in products and internal workflows, which is where enterprises and cloud customers feel the impact. An emergency switch also raises a practical question that a政策 document cannot answer on paper: whether it works when it is needed. That is why the proposal calls for regular testing by independent groups.

What prompted the order, and why the 2028 run matters

Daily Bruin describes the order as a response to recent AI incidents, including one in which 700 AI agents attacked Hugging Face during an OpenAI cybersecurity experiment in July. That account appears in only one of the sources reviewed for this story and has not been confirmed by a second source, so treat it as a reported claim rather than established fact. For background on how incident disclosures are playing out in Washington, read OpenAI discloses four more rogue AI incidents as Congress weighs liability.

This is not Newsom’s first AI executive order. Daily Bruin reports that he has signed four other AI-related orders since 2023, all instructing the state to study the technology.

NBC News describes Newsom as widely seen as a 2028 presidential hopeful and as part of a growing group of Democrats calling for stronger AI regulation. Al Jazeera ties his AI push to that possible run. Newsom’s own statement frames the work around timing: “We’re going to speed up our work on substantial and responsible AI oversight before it’s too late.” Quartz reported the statement and cited The New York Times.

Presidential ambitions can accelerate a governor’s agenda, and they can also make a policy read like a campaign plank. Whether the 2028 framing shapes the pace is an open question. The clearest signal will be whether the November recommendations contain specific audit and shutdown standards or broad goals.

Federal rules are still mostly voluntary

The state move follows a federal shift. On October 9, the White House’s Super Intelligence Force said AI companies must promptly disclose and remedy incidents involving their models, calling it a “not optional” obligation, according to Mezha. Mezha’s headline ties that federal move to disclosures from Anthropic. The federal requirement did not specify penalties or enforcement, and before it, the administration had relied mainly on voluntary frameworks and industry self-regulation.

That gap is where the friction will show. A state shutdown mandate would carry legal force, while the federal incident rule, as reported, has no described enforcement. It is not yet clear how California’s proposals would interact with a federal requirement whose enforcement mechanism has not been explained. Other states are watching too. Massachusetts is weighing independent AI safety tests for frontier models, and any California rule would likely set a template for other states.

Who would feel it, and what to do now

Developers of frontier models face the most direct exposure if the proposals survive. Daily Bruin reports that 32 of the top 50 private AI companies in the world are based in California, which makes the state a practical testing ground for any rule. Enterprises that build with or buy AI agents would feel indirect effects, because a shutdown requirement would reach the tools they run.

None of the steps below are required by the order itself. They are general guidance, not legal advice:

  • Track the mid-November recommendations and any special session call.
  • Review incident-response plans for AI tools: who can pause an agent, who gets alerted, and how logs are kept.
  • Audit access controls so AI agents can reach only the systems they need, and require human approval for high-impact actions.
  • Ask AI vendors how they handle incident reports and how quickly they can disable a model or agent.

The federal incident-reporting guidance points in the same direction. That overlap is the most practical takeaway for now.

What to watch before the deadline

Several open questions will shape the next few weeks. Does the expert panel recommend a mandatory kill switch, and if so, how would the state enforce it? Will the Legislature convene a special session, and when? Do the Hugging Face details hold up under scrutiny? And how will California’s proposals fit alongside a federal incident rule that has not explained its enforcement?

The next milestone is mid-November. If the recommendations arrive with specific audit and shutdown standards, California will have moved from study to proposal, and other states and AI developers will have a concrete draft to react to. If they arrive as broad guidance, the debate will shift to whether a kill switch can be tested in practice, which is a harder question than whether one should exist.