The proposal, in the letter's own structure

Start with who finds the risk. Anthropic does not propose that a regulator test models from outside, after the fact. It proposes independent evaluators working inside frontier AI companies, with access comparable to the companies' own risk staff (per Politico's reporting), reporting what they find to California's attorney general. According to Politico's reporting, developers would pay into a state-administered fund rather than paying the evaluators directly, and state officials could rate and assign them. The proposed funding arrangement is intended to reduce the conflict that can arise when a company directly pays the evaluator assessing it. A state-administered fund could help separate payment from the outcome of an assessment, although the degree of independence would also depend on how evaluators are selected, assigned and protected from interference.

Reuters directly quoted Anthropic's letter as saying: 'When a review finds a model poses a serious risk, the government should be able to ask a court to block it.' Politico's account of the letter's more detailed proposal describes a narrower trigger: a finding of near-term risk of catastrophic harm, after which the attorney general could seek a court order suspending the model's deployment, internal use or further development. Because the full letter has not been published, those details should remain attributed to Politico rather than presented as independently verified statutory language.

According to Politico's account of the letter, the proposed court order could reach three distinct activities: a model's deployment to the public, its use inside the company that built it, and further development. That scope is significant because it would extend beyond preventing a public release to potentially restricting work inside a frontier lab. Frontier labs use their strongest models internally, for coding and research, well before release.

Around that core sit two supporting duties. Large developers would publish risk reports at least every six months. And every frontier company should be able to shut its models down rapidly, whether those models are served to customers or used internally.

The part that cuts the other way

The same letter argues against the tool that gives this debate its shorthand. Requirements that developers build kill switches capable of shutting down entire models, Anthropic wrote, could provide a "false sense of reassurance" to officials. Its reasoning, as reported, is about granularity: most failures will not call for a whole model to go dark. The interventions it backs instead are targeted ones, disabling a problematic AI agent or cutting off a specific customer's access, steps a company can take at the level where the problem appears.

Anthropic is not simply rejecting intervention. Its proposal combines targeted controls, such as disabling a problematic agent or cutting off a customer's access, with a judicial route for cases that meet the more serious risk threshold described in Politico's account. The letter's position is that whole-model shutdown requirements alone could provide false reassurance, not that rapid shutdown capability is unnecessary.

There is context for why a court-based version might appeal to this particular company. In June 2026, the Trump administration ordered Anthropic to immediately restrict foreign access to its cyber-capable Mythos 5 and Fable 5 models on national-security grounds; the Commerce Department later lifted the order, as Reuters reported alongside the letter. The history provides context for why a court-based process might appeal to Anthropic: the company had faced an administrative restriction on foreign access to two models, which the Commerce Department later lifted. A process requiring the attorney general to seek judicial authorization would use a different decision-making route, although the letter does not establish how such a procedure would operate in practice.

OpenAI stopped one step shorter

Anthropic was not the only lab writing to Newsom. OpenAI sent its own recommendations on October 20, and according to Politico's reporting it backed shutdown controls and auditors working inside companies, the two planks it shares with Anthropic, but did not go as far as supporting court orders to block a model.

The two letters share support for embedded evaluators and shutdown capability. The distinction reported by Politico is that Anthropic went further, proposing a route for the state to seek a court order blocking or suspending a risky model. OpenAI's recommendations, as reported, did not include that step. Two letters establish a difference between these companies' proposals, not an industry-wide consensus.

A proposal with a deadline, not a law

None of this is in force. The letter answers Newsom's September executive order (N-9-26, per The Circuitry's reporting of Politico's account), which set state officials to study frontier-model oversight, including embedded auditors and a kill switch, and to report back by November 16, 2026. What California does with two labs' letters and that report is still open. Newsom's office did not immediately respond to Reuters about Anthropic's letter.

California is also not the only venue testing the idea. Reuters notes that a bipartisan group of US senators was debating legislation in September that would let the federal government block the release of models deemed unsafe, with companies able to challenge the decision in federal court; Congress left for its pre-election break and the proposal has not moved. House lawmakers in July proposed an AI Kill Switch Act that would let officials order shutdowns of models putting human life or the economy at risk. Both remain proposals.

So the letter's practical weight is narrower than its headline. It does not create a power, set a standard, or name a model. What the letter does is put a significant version of the proposed oversight power on the table, from a company that could itself be subject to it, ahead of the governor's working group's November 16 deadline. Whether California adopts that approach, modifies it or rejects it remains open.

Sources

  1. [1] The full October 15 letter was not publicly available. Reuters independently confirmed its authenticity with Anthropic and directly quoted one sentence; the more detailed provisions and the OpenAI comparison are attributed to Politico's reporting.
  2. [2] Reuters (via Duke FM syndication), "Anthropic tells California governor courts should be able to block risky AI models," October 28, 2026 — the version read: the letter's date and confirmed authenticity, the directly quoted sentence, the kill-switch warning and targeted interventions, rapid shutdown capability, the June 2026 Mythos 5 and Fable 5 restriction and its lifting, Newsom's September panel and his office not immediately responding, and the federal proposalsRead source
  3. [3] The Circuitry, "Anthropic Tells Newsom Courts Should Be Able to Halt Risky AI Models" (Politico reporting), October 28, 2026 — the detailed proposal as reported by Politico, including the evaluator fund and the executive order N-9-26 framingRead source
  4. [4] Daily AI Brief, "Anthropic backs court power to block risky AI models" (Politico reporting), October 28, 2026 — the evaluator fund, state officials rating and assigning evaluators, and access comparable to internal risk staffRead source
  5. [5] USA News Hub, "Anthropic tells Governor Gavin Newsom courts should be able to block risky AI models" (Politico reporting), October 29, 2026 — corroborating carrier of Politico's account, including the OpenAI comparison and the November 16, 2026 working-group deadlineRead source