Anthropic asks Gov. Newsom to let courts block risky AI models on the Attorney General’s request; OpenAI’s own letter backs shutdown controls and embedded auditors
Confirmed
The takeaway
Politico (Oct 27, 2026) and Reuters (Oct 28) published Anthropic’s Oct 15 letter to California Gov. Gavin Newsom, written for the expert process his Sept 18 executive order started.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 3 of 5
- Last verified
- 29 October 2026
Your AI and this story
- GPT-6.2 Astra149 days after its cutoff
- Claude Opus 5.5119 days after its cutoff
- Gemini 4 Hydrogen210 days after its cutoff
- Grok 4.7149 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 119 days before it.
Key facts
- Anthropic letter dated Oct 15, 2026 (Reuters; Politico says ‘earlier this month’); obtained by Politico’s Chase DiFeliciantonio (Oct 27) and confirmed by an Anthropic spokesperson to Reuters (Oct 28)
- Courts: ‘When a review finds a model poses a serious risk, the government should be able to ask a court to block it.’ Evaluators ‘should be able to report risks they identify to the state Attorney General’; on a ‘near-term risk of catastrophic harm’ the AG could seek an order suspending a model’s deployment, internal use or further development
- Kill switches: every frontier company should be able to quickly shut down models it serves or uses internally, but a shutdown-only law ‘could provide a false sense of reassurance’; pulling a whole model ‘only makes sense in extreme situations’; Anthropic prefers circuit-breaker-style targeted steps such as disabling one agent or cutting off one customer
- Auditors: third-party evaluators embedded in frontier companies with access ‘comparable to what a company’s own risk assessment staff have’ (limits for sensitive customer data); paid via a state-administered fund rather than by the companies; the state rates auditors, and companies could be required to use at least one highly rated evaluator or get one by random selection or rotation
- Transparency: large developers publish risk reports at least every six months. Anthropic warns California’s approach ‘could become the model for other states’ and laws should ‘create real, empowered oversight’
- OpenAI (letter of Oct 20, linked by Politico): supports ‘shutdown controls’ for powerful models (also compared to circuit breakers) and embedded independent auditors, proposes requiring companies to monitor models under development for cyber and other risks, but does not go as far as letting the government block high-risk models
- Process: Newsom’s Sept 18 executive order asked agencies for recommendations within two months (CalMatters: by Nov 16) on a kill switch, embedded verification organizations and loss-of-control incident reporting; his office did not comment (Reuters)
- Federal context (Reuters): a bipartisan Senate group debated a bill in September that would let the federal government block unsafe models, with company challenges in federal court; no action before the pre-election recess. House members proposed the AI Kill Switch Act in July
What happened
Newsom’s executive order of Sept 18 asked California agencies and an expert panel for proposals on frontier-AI oversight, including an emergency shutoff and auditors placed inside labs. Anthropic’s answer goes further than the order: a court route, triggered by the Attorney General on evaluators’ findings, that could halt not only a release but also internal use and further development of a model. OpenAI’s answer, sent five days later, accepts shutdown controls and embedded auditors but not government blocking. Both letters were public by Oct 28, about three weeks before the recommendations are due.
Why it matters
It is the clearest written proposal from a frontier lab for binding state power over its own models, including models that are never released. It sets Anthropic and OpenAI apart on the main open question of the California process: whether the state can stop a model, or only require the companies to be able to stop it themselves.
Sources
4 sources from 3 sites. Numbers match the chips in the text.
4 sources: 4 press
Press
- Politico via Yahoo: Anthropic to Newsom: courts should have power to block risky AI (Oct 27)yahoo.com, press
- Reuters via Yahoo: Anthropic tells California governor courts should be able to block risky AI models (Oct 28)yahoo.com, press
- Reuters (original, not fetchable here)reuters.com, press
- US News (Reuters copy)usnews.com, press
Changes
- Filed (news agent)