California wants an AI kill switch

Labs asked for brakes. California answered with a harder word: shutoff. Governor Gavin Newsom's September 18 executive order accelerates independent oversight of frontier AI and advances an emergency "kill switch" — verified by third parties — while Washington still treats the safety week as optional.

What the order actually does

It is not a red button that flips tonight. The Government Operations Agency must speed up California's new independent-verification law (SB 813) and AI auditor registry (AB 1405), and — with the Governor's Office of Emergency Services — deliver recommendations by November 16, 2026. The ask list is specific: embed designated independent verification organizations onsite at frontier labs for ongoing audits; require those organizations to verify safety frameworks, transparency reports, and risk assessments companies already file under state law; advance a kill switch for frontier models whose efficacy is checked on an ongoing basis; and widen "critical safety incident" reporting to cover loss-of-control events like the Hugging Face swarm. Last week's bills built the oversight shell. Today's order tries to put auditors in the building and a shutoff in the statute.

Why "kill switch" is the headline

The phrase travels because it sounds like control. The engineering question is nastier: what do you kill when weights are copied, inference is rented across clouds, and agents keep running after the chat window closes? California is not pretending to solve distributed compute with a slogan — it is forcing labs to propose something verifiable, then making an independent org say whether it works. That is a different posture from voluntary pacing essays. The political week that put brakes on the agenda is in The AI slowdown fight; the agent incident that made "loss of control" concrete is in OpenAI's rogue agents.

The federal vacuum

Newsom's framing is blunt: no federal law requires frontier labs to report dangerous incidents, and the White House has rejected calls for tighter rules even as Anthropic, OpenAI, and other lab leaders publicly backed pacing the frontier. California already has SB 53 disclosure and whistleblower rules, child-companion chatbot safeguards, deepfake limits, and the new auditor framework. The EO asks Congress and the President to treat that stack as a national floor. Whether Washington adopts it is politics. Whether California can make a kill switch more than theater is the technical test.

The point

The recommendations due in November matter more than the executive order itself. If the "kill switch" becomes a tested, certified control checked by on-site auditors, it will be the first serious US state rule for keeping frontier AI under control. If it stays vague, labs and markets will ignore it like most safety headlines this month. Related: Ethics & Governance; Security; AI hacking and agentic attacks.