A blocking power, with the bar set first
Anthropic has spent this year arguing in public that governments should be able to stop a dangerous model release; its Advanced AI Framework says so in plain terms. What is new in the Politico reporting, summarised by Reuters on October 16, is the translation of that position into legislative language — and the numbers the translation carries. The draft text itself is not public, and everything about it here is as reported.
As reported, the blocking test has two parts, and both have to be met. The consequence has to clear one of two lines: a foreseeable, imminent risk of more than $1 billion in damages, or of killing or seriously injuring more than 50 people. And the risk has to stem from one of three enumerated sources: the creation or release of a biological weapon, a cyberattack, or the model evading its developer's controls. A model that is merely very capable, misused in smaller ways, or risky on a longer horizon does not meet the test as described.
The high bar is deliberate, and Anthropic's own framework page presents it that way — as a guard against government overreach. Read from the other side, the same bar decides in advance which fights the law will never have. In our assessment, few contested launches have even alleged harm approaching that scale, so a regime drawn this way would leave most real-world disputes about model releases outside its reach.
One difference between the draft as reported and the framework behind it is worth noting. Anthropic's framework lists four catastrophic-risk categories — biological, cyber, loss of control, and automated R&D, the fourth amplifying the other three. The draft's blocking test, as reported, names three sources; automated R&D is not among them.
Who watches, and who pays them
The checking machinery, as reported, sits with independent evaluators. They would register with the Commerce Department and periodically evaluate both the lab and its frontier models. Where evaluators determine an "imminent enumerated risk" — the draft's defined test, not any concern an evaluator happens to notice — they must report it to Commerce within 72 hours of completing their assessment.
Labs could pay the evaluators who assess them, provided the payment is not tied to the evaluation's outcome. That clause deserves a plain reading: outcome-independent pay does not by itself guarantee independence. Payment that is formally detached from any single result can still shape who gets hired again, which is the argument that played out in public this month when labs endorsed independent review while negotiating the terms around it. Registration with Commerce is the draft's partial answer: the state keeps the list, even if the labs keep the invoices.
One door to a halt, and it is a courtroom
If a halt happens at all, it happens in one place. As reported, Commerce could ask a federal judge for the District Court for the District of Columbia to order a model suspended. The draft then closes every other door: its language prohibits any other US government department, agency or officer from suspending a model's development, deployment or internal use.
That clause is the most consequential line in the pitch. It does not just create a power; it centralises it, and in doing so it removes the improvisation route, where an agency reaches for whatever authority it already holds. Labs that have complained about patchwork enforcement get predictability. Regulators who wanted a faster lever than a court filing do not get one.
The scope is a short list — and it is the framework's
The spokesperson's link to the Advanced AI Framework matters because the framework is where the scope figures live; they are Anthropic's framework numbers, which the draft is said to reflect, not text quoted from the draft itself. On Anthropic's own page, the framework's rules would apply only to models trained using more than 10^25 FLOPs, developed by companies earning more than $500 million in AI-related revenue or spending more than $1 billion on AI R&D. Those two company tests are alternatives, not cumulative gates: either line is enough. The framework's developer duties run alongside: test models, publish findings, safety frameworks, system cards and regular risk reports, engage independent evaluators, and maintain a security program.
That scope is a handful of companies. Everyone below those lines sits outside the regime entirely, including labs whose models are weaker on paper and less carefully evaluated in practice. There is a case for that — catastrophic risk, if it exists, sits with the largest training runs. But a law drawn around the current frontier also freezes a picture of where risk lives, and thresholds of this kind are ones a well-advised company can plan around.
What to watch
None of this is law. It is a draft, pitched privately, with no named sponsor in the reporting and no public text to check the summaries against; Commerce did not immediately respond to Reuters. The next facts that matter are simple: whether a lawmaker attaches a name to it, whether any text that surfaces matches the thresholds reported, and how OpenAI and Google — which would live under the same lines — respond to a rival writing the first draft of their regulator's rulebook.
Sources
- [1] Reuters, summary of Politico's reporting, October 16, 2026 (syndicated copy) — the pitch, the blocking thresholds, the evaluator and court machinery, all as reportedRead source
- [2] Anthropic, Policy on the AI Exponential — the Advanced AI Framework on Anthropic's own page: scope, risk categories and developer dutiesRead source
- [3] Anthropic, AI policy pageRead source