California's New AI Kill Switch Order, Explained

On September 18, 2026, Governor Gavin Newsom signed an executive order directing California's Government Operations Agency to convene outside experts and recommend, within two months, ways to strengthen the state's AI safety and security laws. Among the specific measures the order asks those experts to consider: an emergency kill switch for the most advanced AI systems, independent verification of frontier AI companies' safety frameworks, and expanded reporting requirements for loss of control incidents.

Nothing in the order changes what a small AI builder has to do this week. It is a directive to study and recommend, not a new statute. Here is what it actually says, how it connects to the laws already on the books, and what is worth watching over the next two months.

What the order directs

The order tasks the Government Operations Agency, working with the Governor's Office of Emergency Services, with convening national experts on AI safety and bringing back recommendations within 60 days. Newsom framed the urgency plainly:

“We're not waiting to act, we're going to speed up our work on substantial and responsible AI oversight before it's too late,” the governor said in the announcement, published on the state's official site.

That expert panel is being asked to look at four specific ideas: independent verification organizations embedded on-site at frontier AI companies to run regular audits, mandatory third-party verification of the safety frameworks and transparency reports those companies already publish, an AI kill switch for frontier models with ongoing checks that it actually works, and a broader definition of what counts as a reportable safety incident, explicitly including loss-of-control events like the recent Hugging Face attack.

This builds on laws that already exist

The order does not arrive in a vacuum. It leans on two pieces of legislation Newsom already signed: SB 813, which sets up independent verification frameworks, and AB 1405, which creates a registry of approved AI auditors. Both are supply-side laws. They create the auditors and the verification machinery; they do not, by themselves, impose a new audit duty on a small business using AI. We covered exactly that distinction, and where the real procurement exposure sits for small builders, in our breakdown of California's AI audit laws.

The order also references SB 53, the nation's first AI safety law requiring frontier labs to disclose safety practices and report incidents, as the baseline this new push is meant to strengthen.

The kill switch idea itself

A kill switch, in this context, is not a single red button. It is the ability to halt or roll back a deployed model's behavior when something goes wrong, at whatever layer that requires: revoking API access, disabling a specific capability, or rolling back a model version entirely. We wrote a longer explainer on what an AI kill switch actually needs to work, covering the three layers most proposals collapse into one. Today's order does not specify implementation. It asks the expert panel to figure out what “works” even means, including how you would verify a kill switch stays effective as a model is updated.

What this means if you build on AI, not build AI

If your business uses AI, rather than trains frontier models, this order is aimed several rungs up the ladder from you. The verification and kill switch requirements under discussion target the handful of labs building the most capable systems. Nothing here creates a new compliance checklist for a small SaaS company calling an API.

What is worth tracking: the two-month clock the order sets runs out around mid-November 2026. Whatever the expert panel recommends will likely shape the next legislative session, and past precedent (SB 53's incident reporting rules) has a way of showing up indirectly, in the terms-of-service and status pages of the model providers you depend on. Watching for changes there is a better use of your time this quarter than trying to read the executive order as a compliance document, because right now it is not one.

For the wider context: our AI safety and risk hub tracks every one of these stories as they land, and our guide to keeping up with AI news without doomscrolling covers how to filter signal like this from the noise.

The bottom line

Newsom's order is a research-and-recommend mandate, not a new law. It formalizes California's intent to pursue kill switch requirements and independent verification for frontier AI companies, sets a two-month deadline for expert recommendations, and explicitly widens the definition of a reportable AI safety incident. For everyone building on top of frontier models rather than inside them, the honest answer is: watch, do not panic, and expect the real detail in November.