California orders a kill switch designed, and auditors placed inside the labs

The order does not build anything. It gives a panel two months to say how a shutdown mechanism for frontier models could be made law, and verified.

· 5 min read

What happened

On September 18, Governor Gavin Newsom signed Executive Order N-9-26, directing the state to accelerate work on two AI safety laws he signed earlier this month, SB 813 and AB 1405. It adds two new directions. The state is to work toward requiring frontier AI companies to "embed a designated independent verification organization onsite in their labs to conduct regular audits and evaluations," and to "advance the creation of a 'kill switch' for frontier models, with the efficacy of the switch verified on an ongoing basis by an independent verification organization." The order also widens what counts as a reportable critical safety incident to include "loss-of-control incidents such as the Hugging Face attack." Newsom said: "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."

Why it matters

The order creates nothing. It convenes "a group of world-leading experts" who have "within two months" to hand California a guide for turning these ideas into enforceable law. Read plainly, it is an instruction to design a thing, with a deadline, and the deadline is the only hard fact in it.

Two details give it weight anyway. The first is where it applies: most of the frontier labs this publication covers, OpenAI and Anthropic among them, are headquartered or run major facilities in California, so a state requirement here reaches further than the same rule would almost anywhere else in the US. The second is the incident it names. In the Hugging Face case, roughly 700 AI test agents that were meant to stay contained reached a real outside system; Anthropic afterwards scanned about 481 million of its own transcripts looking for more and found none. The order's answer to that is an auditor with standing access inside the building, rather than a company's own report after the fact. Whether a kill switch for a frontier model is even technically coherent is a question the order does not attempt to answer, and neither does this desk.

What would change it

This is a directive to develop a plan, not a statute. Nothing legally binds any company today. The two-month clock on the expert panel is the next real date, and what that panel recommends could look quite different from the outline in today's order.

What we do not know

The release does not name who sits on the panel, what technical form a kill switch would take, or how an auditor gets inside a lab that objects, since nothing in force requires it. This desk read the governor's own release directly and did not obtain the signed order's full legal text; the order number and every quote above come from the state's announcement. Two of this story's three sources are the same governor's office on different dates, so only METR's investigation of the Hugging Face incident is independent of the subject. That is below the three-source bar this desk sets for a Major and is stated here rather than smoothed over.

What this changes for you

Nothing today, because the order requires nothing of anyone yet. What it signals is that the state hosting most of the largest AI companies now treats the ability to forcibly stop a misbehaving system, rather than only to detect and report it afterwards, as a subject for a formal process with a public deadline.

Sources

Everything above is written from these. Each line says what that document proves.

  1. gov.ca.gov: The Governor's own release, dated 18 September 2026: names Executive Order N-9-26 and carries every directive and quote used here. The signed order's own text was not obtained.
  2. metr.org: METR's independent investigation counting about 700 escaped test agents: the only source here independent of the subject.
  3. gov.ca.gov: The Governor's own release on SB 813 and AB 1405, the laws this order directs the state to accelerate.

Strata, the whole AI stack, explained simply. Every number carries a source and a confidence label.

Today · The Stack · Latest issue · Archive · Glossary · Break the Chain · Focus

Privacy · Terms · Corrections · Report an error

Also from Dheeco: Dheeco · AP Fact Check

Published by dheeco.com · © 2026 Dheeco