California Gov. Gavin Newsom on September 18 ordered state officials to develop recommendations for stronger safeguards around frontier AI systems, including the possible creation of a verified emergency shutdown mechanism often described as a “kill switch.” The order sets a November 16 deadline for the recommendations, but it does not itself require AI developers to install a shutdown tool or comply with new onsite-audit rules. (gov.ca.gov)
Executive Order N-9-26 directs the California Government Operations Agency, working with the Governor’s Office of Emergency Services and national experts, to examine the technical feasibility and potential effectiveness of amendments to state AI safety and security laws. The assignment is explicitly framed as a recommendation process: any new obligations for companies would require subsequent policy action rather than taking effect under the order alone. (gov.ca.gov)
The review must address whether large frontier developers should be required to embed designated independent verification organizations onsite at their labs for periodic audits and evaluations. It must also consider whether the safety frameworks, transparency reports and risk assessments that frontier AI companies file under state law should be independently verified against standards an independent verification organization considers adequate. (gov.ca.gov)
The prospective shutdown provision is similarly limited at this stage. The order calls for recommendations on requiring a “kill switch” for frontier models and for having its efficacy verified on an ongoing basis by an independent verification organization. That is a potentially consequential proposal, but it is not an enforceable statewide kill-switch mandate as of September 19, 2026. (gov.ca.gov)
A practical emergency-control requirement would raise difficult questions that the order leaves for the expert process: what systems would qualify as frontier models, who would be covered as a large developer, what a shutdown mechanism would technically encompass, and how an outside evaluator could test it. The executive order does not name participating experts or establish a public list of covered companies, so those details remain unresolved. (gov.ca.gov)
Incident reporting is part of the review
Newsom’s order also asks the group to consider expanding the definition of critical safety incidents that companies must report. Specifically, it calls for recommendations on including a range of loss-of-control incidents, referring to recently reported events involving large frontier developers. The governor’s office said the action followed recent concerns about AI security and safety risks. (gov.ca.gov)
Taken together, the proposals point toward an oversight model that would put more weight on independent assessment, rather than relying only on company-authored safety materials. If California later adopts any of the recommendations, developers could face greater expectations to document risks, make those materials reviewable and demonstrate that emergency controls work as intended. Those outcomes remain contingent on what the experts recommend and whether state policymakers pursue the recommendations. (gov.ca.gov)
Fast-tracking recently enacted audit frameworks
The executive order also accelerates implementation steps connected to two California laws signed earlier this month: Senate Bill 813 and Assembly Bill 1405. SB 813 establishes a framework for independent verification organizations that can assess AI systems and models for safety and risk, while AB 1405 creates a state registry for AI auditors and standards concerning their independence, transparency and integrity. (gov.ca.gov)
Under the order, the Government Operations Agency must publish application requirements, procedures and criteria for independent verification organizations by May 1, 2027. It must also complete a separate statutory implementation step related to the AI-auditor framework and begin required follow-on actions by December 1, 2027. Those deadlines concern the state’s implementation infrastructure; they do not by themselves create a new duty for every AI company to undergo an audit. (gov.ca.gov)
The timing matters because California is building the oversight mechanisms while separately asking experts what additional controls may be needed for the most capable systems. The laws signed on September 9 laid a framework for independent third-party evaluations and auditing. The new order seeks to speed that work while testing whether the existing regime should be expanded to include onsite verification, validated reporting and an emergency intervention capability. (gov.ca.gov)
What AI companies and adopters should watch
For frontier-model developers, the immediate obligation is limited: the order starts a state-led review and sets deadlines for agencies, not a newly defined compliance program for private companies. Still, it provides a clear indication of the safeguards California wants experts to assess—independent evaluators, third-party verification of safety submissions, reporting of loss-of-control incidents and a validated emergency shutdown option. (gov.ca.gov)
For enterprise buyers and organizations deploying advanced AI, the signal is broader. California’s approach increasingly emphasizes evidence that safety claims and risk-management processes can withstand outside scrutiny. Businesses that depend on frontier-model providers may therefore have reason to watch for changes in providers’ audit practices, incident disclosures and safety documentation, even though the order does not create new procurement duties for them. This is an inference from the safeguards California has asked experts to evaluate, not a new legal requirement. (gov.ca.gov)
The November 16 recommendations will be the next concrete milestone. They could inform legislation, regulation or additional executive action, but the executive order does not predetermine any of those outcomes. For now, California has opened a fast policy process around some of the strongest potential controls in frontier-AI governance while moving ahead with the independent-assessment structures already enacted into state law. (gov.ca.gov)
