Newsom orders a kill switch and onsite auditors for frontier AI
California Gov. Gavin Newsom signed an executive order Friday that tells his own agencies to design the state's next round of AI rules — and names the two things the industry has resisted hardest: a kill switch for frontier models, and independent auditors embedded inside the labs.
Executive Order N-9-26 gives the Government Operations Agency until November 16, 2026, to deliver recommendations on four specific mandates, and each one aims at the same target. Frontier developers would have to host independent verification organizations onsite to run periodic audits and evaluations. Their safety frameworks, transparency reports and risk assessments would have to be independently verified rather than simply filed. A kill switch for frontier models would have to exist, with its efficacy checked on an ongoing basis by a verifier. And the definition of a reportable critical safety incident would widen to cover loss-of-control events — the kind of behavior major labs have spent the past weeks disclosing.
The order also puts hard dates on the verification machinery California already passed. Application criteria for independent verification organizations must be posted by May 1, 2027, and the state must begin acting under its AI audit statute by December 1, 2027. That matters more than the four asks: those are recommendations, these are deadlines the agencies have to hit whether or not the Legislature moves. We covered the certification framework when it passed — California just made AI companies face an auditor they don't pick.
The political framing is explicit. In the order's preamble, Newsom cites "multiple instances of apparent attempts" to use AI to build bioweapons and of agents working, sometimes collectively, to defeat labs' own security protocols — and says California will keep leading "absent federal action — which unfortunately is not forthcoming due to a failure of leadership by the President and Congressional leaders." He is also picking up his own thread from the week before: this is the executive action he floated when he told POLITICO he was weighing a special session, an order, or agency rulemaking instead of a bill, as we wrote in Newsom floats special session or executive action on AI safety.
The read: the deadlines are the teeth. A kill switch requirement cannot be waived by a legislature that never votes, and an onsite auditor, once admitted, is not a self-report. What California cannot do is make the labs build it — nothing in N-9-26 compels a single developer today. It commits the state to showing up in 2027 with rules already written.
xAI released Grok Voice Transcribe 2.0, a speech-to-text model it says is twice as accurate as its predecessor at identical pricing — $0.10 per hour of audio in batch mode and $0.20 per hour streaming, with diarization, word-level timestamps and key-term prompting included. That flat price is the story in a category where accuracy improvements have usually arrived with a rate card change.
The company says the model ranks first for accuracy among 32 streaming models on Artificial Analysis's leaderboard. The specifics it published are blunter than the summary: on short in-car-style commands, word error rate drops from 20.6% to 6.8%, and on telephony audio from customer-support calls it claims to lead every model tested. Multilingual switching mid-recording in a single pass is the largest single improvement. It is a model behind the API, not open weights — the Hacker News thread's first reaction was exactly that — and xAI says existing API integrations pick up the accuracy gain with no code changes.
If a state can mandate a kill switch but cannot force a lab to build one, what would actually give the requirement teeth? Tell us in the comments.
