California issued an executive order on Friday (18) to speed up independent oversight of artificial intelligence systems. It also requests recommendations for advancing an emergency shutdown mechanism, often called a “kill switch,” for frontier models.
The verb matters: California has not made such a mechanism mandatory. The order directs the Government Operations Agency, working with the state's emergency services office, to convene national experts and deliver recommendations within two months. One proposal they must consider is ongoing independent verification that the shutdown mechanism actually works.
Auditing outside the lab
The action accelerates two recent state laws. SB 813 created a framework for certifying independent organizations qualified to assess model safety and risk. AB 1405 established a registry for AI auditors and standards for independence, transparency and integrity.
The order now asks the expert group to consider more concrete requirements: external verifiers embedded in frontier AI labs, independent validation of safety plans and risk reports, and an expanded definition of critical incidents that includes loss-of-control events.
That shifts the center of gravity in governance. Instead of accepting only the developer's own report, the proposal brings auditing closer to the actual model development and operating cycle. The logic will be familiar to anyone working with regulated software: important controls need to be testable, repeatable and separated from the team shipping the system.
A button is not a complete system
“Kill switch” is a powerful visual phrase, but it compresses a hard technical problem. Models may run across regions, operate through agents and connect to third-party services. Turning off one endpoint does not necessarily stop copies, already delegated tasks or credentials that remain active.
That is why the least cinematic part of the order may be the most useful: continuously verifying that the mechanism works. An emergency control is real only when it comes with a dependency inventory, credential isolation, audit trails, recurring tests and clear authority to activate it. Without those pieces, the big red button is expensive decoration.
What this means for AI companies
The recommendations still need to be developed and may require further changes to state law. Even so, the direction is clear. Companies building frontier models in California could face an additional layer of independent verification over risk assessments, safety plans and incident response.
The discussion also matters to teams consuming these models. Continuity contracts, fallback paths and the ability to switch providers are no longer only cost questions. If a critical model must be interrupted, the customer's architecture needs to degrade safely.
California is trying to turn a simple metaphor into a verifiable control. The real test will not be building the button. It will be proving that the entire chain knows what happens when someone needs to press it.