California orders an AI kill switch for frontier models
Executive Order N-9-26 sets deadlines for independent AI audits and asks experts to design a kill switch for frontier models, though nothing has to be built yet.
California is not waiting for Washington. On 18 September Governor Gavin Newsom signed Executive Order N-9-26, directing state agencies to accelerate independent oversight of artificial intelligence and to work out what a mandatory emergency shutdown for frontier models would actually require.
What the order actually does
Two of the deadlines are concrete. The Government Operations Agency must complete the requirements of Government Code Section 8898.1 and publicly post the application requirements, procedures and criteria for independent verification organisations by 1 May 2027. It must complete the requirements of Section 11549.82 by 1 December 2027.
A third deadline is where the headline lives. By 16 November 2026 the agency, working with the Governor's Office of Emergency Services and outside experts, must send Newsom recommendations on amendments to state law covering AI safety and security.
The four things on the list
First, requiring large frontier developers to embed designated independent verification organisations onsite in their labs to conduct periodic audits and evaluations. Second, independently verifying the safety frameworks, transparency reports and risk assessments that frontier AI companies already have to file. Third, requiring the creation of a kill switch for frontier models, with its efficacy verified on an ongoing basis by an independent verification organisation. Fourth, widening the definition of a critical safety incident to include a range of loss-of-control incidents.
Why now
The order cites recent alarming incidents, naming the Hugging Face attack, and argues that AI chief executives themselves have asked for rules. It also leans on work California has just finished. SB 813 (McNerney), signed the previous week, establishes the framework for independent verification organisations, while AB 1405 (Bauer-Kahan) creates a state registry for AI auditors and sets standards for their independence.
Our opinion
Read the operative paragraphs and the gap is obvious: this order does not switch anything off, and it is not built to. Its deadlines produce requirements, recruitment criteria and a recommendation list, which is why the phrase kill switch is doing a great deal of political work in a document that mostly asks for homework due next year. That is not a reason to dismiss it. Independent verification only bites once somebody writes the standards, and California is at least scheduling that work while Congress does nothing at all. The credibility test now has a date on it. If 16 November produces a recommendation list indistinguishable from what the labs would happily write themselves, audits performed by firms the labs pay for, then the strongest AI safety framework in America will have amounted to a timetable. The onsite-auditor proposal is the one to watch, because it is the only item on the list the companies cannot easily absorb into business as usual.