Hidden Anthropic bill: high threshold, single legal channel and preemption of state AI regulation
The draft, which Politico obtained and Anthropic itself confirms, would halt a model only for imminent risk of over $1bn in damages or more than 50 killed or seriously injured — and only from three named sources of harm. Administrative power would be concentrated in the Commerce Department and a single federal judge in Washington, DC, while state laws on risk disclosure and external evaluation are overridden.
The story: a confidential draft from the lab itself
Representatives of AI company Anthropic have quietly presented lawmakers on Capitol Hill with a draft bill to regulate the safety of frontier models — the most advanced AI systems. Politico reported this, citing three anonymous sources familiar with the private discussions. The draft, which according to Politico was not previously publicly known, was reported on October 16, 2026 — days before Congress recesses ahead of the November elections.
An Anthropic spokesperson confirmed to Politico that the draft reflects the proposals in the company's own "Advanced AI Framework": "This bill represents the proposals in Anthropic's Advanced AI Framework, which we have been advocating to federal and state legislators since June," the spokesperson said.
It is this dual role that makes the story remarkable: a commercial frontier lab has drafted legislation meant to govern its own technology, while Congress remains divided over how the industry should be regulated at all.
What the draft would actually do
The threshold for intervention is very high. The government could halt a model's deployment or use only if it posed a foreseeable, imminent risk of more than $1bn in damages, or of killing or seriously injuring more than 50 people. And the risk must stem from one of three specified sources: creation or release of biological weapons, the model evading its developer's controls, or a cyberattack. Politico reported that the language mirrors California state legislation, including SB 53, which was signed into law last year.
The enforcement path is narrower still than the threshold. Only the Commerce Department could seek a halt, and only a judge of the US District Court for the District of Columbia could order a model suspended. The draft expressly bars all other departments, agencies and officials from stopping a model's development, deployment or internal use (The News, drawing on Politico's reporting).
The draft also establishes an evaluation regime: independent evaluators, registered with the Commerce Department, would periodically assess the lab and its frontier models, Reuters reported, citing Politico. Labs could pay the evaluators — but only if the payment is not tied to the outcome of the evaluation. Evaluators are required to report "imminent enumerated risk" to the department within 72 hours. The Commerce Department did not immediately respond to Reuters' request for comment.
The state-power change — and the political headwinds
The draft would also override certain state laws, including those in California, in the areas of risk disclosure, external evaluation and incident reporting. That marks a notable shift from Anthropic's position in June, when the company argued that states should retain authority absent a strong federal law (AFP via The Straits Times). Preemption of state law is precisely what Democratic lawmakers have opposed in earlier proposals, Politico noted.
The draft competes with other proposals: the FRONTIER Act in the House (introduced by Trahan and Obernolte) and a Senate draft associated with Thune, Klobuchar and Cruz. No members of Congress are identified as sponsors of Anthropic's draft, and the bill's fate is unknown — not least because Congress is in recess until after the November elections.
The measured praise — and the criticism
Outside commentators are partly positive. Charlie Bullock of the Institute for Law & AI called the proposal "pretty reasonable" and said it is "more than other frontier labs have done" (AFP/Straits Times). But he also noted it is weaker than the FRONTIER Act in the House.
It is worth being precise about what that comparison means. The FRONTIER Act is a bill written by others; Anthropic's draft is effectively industry's own answer. For a lab to put forward binding, legally enforceable proposals at all — with a named agency, a named court and a 72-hour reporting duty — is more than some competitors have done. At the same time, the draft by design restricts who can act (only Commerce), how they can act (through one court) and when they can act (only for the three listed harm types, above very high thresholds).
What the design choices mean
The three choices underpinning the draft — high thresholds, a narrow legal channel, and state-law preemption — point in the same direction: narrow federal intervention in place of a broad oversight regime.
The threshold of $1bn or 50 dead and seriously injured, tied exclusively to bioweapons, model escape or cyberattacks, leaves large areas outside federal reach. Problems such as bias, deceptive content, labor-market effects, or misuse that does not fit the three categories would not trigger legal grounds to halt a model. And even if the imminent risk is real, Commerce must choose to act, and a single judge in DC must agree.
The preemption also shifts power from the states to the federal level — but only in areas where the draft gives federal authorities tools at all. States that today can require risk disclosure, external evaluation or incident reporting lose that tool, without the draft clearly replacing it with anything stricter.
It is this combination that makes the draft a test: it is at once the most concrete regulatory proposal any frontier lab has put forward, and a bill that — if passed in this form — would leave the government with very few levers over a dangerous model that does not fit the three categories.
What we don't know
Several key matters remain unresolved. The draft's full text has not been made public; everything we know is based on Politico's copy and its three anonymous sources. No members of Congress are openly tied to the proposal. The Commerce Department has not commented, so the administration's position is unknown. And with elections in November and Congress in recess, it is impossible to say whether the draft will be introduced as legislation at all — or whether it primarily serves as a statement of how strict regulation the lab itself considers acceptable.