Anthropic has been quietly circulating a draft AI bill among members of Congress that would decide when the US government may stop a frontier AI model from being released or used. The answer it proposes is narrow: only when the danger is imminent and very large, and only with a federal judge's order.
Politico first reported the draft on 16 October, citing three people familiar with it. Anthropic confirmed to AFP the next day that it was sharing the text with lawmakers. A company spokesperson told Benzinga: "These are sensible policies we believe Congress should act on."
What the Anthropic draft AI bill would do
According to Politico's account, summarised by Business Today and Benzinga, the draft rests on a few core mechanisms.
- A high bar for intervention. The government could block a model's release or deployment only if it posed a foreseeable and imminent risk of causing more than $1 billion in damage, or of killing or seriously injuring more than 50 people.
- Three named risk types. The danger has to come from a cyberattack, the creation of a biological weapon, or a model evading its developer's control. Politico noted that this language tracks California's SB 53.
- A judge, not an agency, decides. The Commerce Department could not shut a model down on its own. It would have to ask the US District Court for the District of Columbia to order a suspension. No other federal agency or official could halt a model's development, deployment or internal use.
- Registered independent evaluators. Auditors registered with Commerce would periodically assess frontier labs and their models. Labs could pay the auditors directly, but pay could not depend on the result.
- A 72-hour reporting window. An evaluator that finds an "imminent enumerated risk" would have to tell Commerce within 72 hours.
Developers would also have to publish safety frameworks and transparency reports for new models, keep whistleblower protections in place, and report critical safety incidents to Commerce, according to Business Today.
The preemption question
The most contested part is what the draft would do to state law. Politico reported that it would override state and local rules on public risk disclosures, third-party audits and critical incident reporting.
That puts Anthropic on a different footing from its earlier stance. In June the company wrote that "absent a strong federal law, states should retain full authority to legislate," AFP reported. Supporters of the new draft would argue it is that strong federal law. Critics, mostly Democrats who have defended rules passed in California, Illinois and New York, have resisted federal preemption in other bills, Business Today noted.
How it compares with bills already in Congress
The draft is not the only federal proposal on the table. In the House, Representatives Lori Trahan (D-Massachusetts) and Jay Obernolte (R-California) have introduced the FRONTIER Act, which would also let the government stop deployment of risky models. In the Senate, Majority Leader John Thune and Senators Amy Klobuchar and Ted Cruz are negotiating a bill on catastrophic AI risks, Benzinga reported.
Charlie Bullock of the Institute for Law & AI called Anthropic's proposal "pretty reasonable" and "more than other frontier labs have done," according to AFP. On X, he said he considered it weaker than the FRONTIER Act.
Why Anthropic wants limits on shutdown power
The timing matters. Benzinga and Business Today both tie the draft to US export controls that took two of Anthropic's most advanced models offline earlier this year. Business Today said the controls were imposed during a series of phone calls, which raised concern in Silicon Valley about arbitrary intervention. AFP separately reported that Washington forced Anthropic to suspend its most advanced models for nearly three weeks in June, and that the Trump administration barred the company from Pentagon work in February.
Read against that history, the draft does two things at once. It accepts that the government should be able to stop a dangerous model. It also tries to make sure that power runs through one agency, one court and one written threshold, rather than through informal pressure.
What happens next
Nothing moves quickly. Congress has not passed a federal AI law, and lawmakers are out of session until after next month's midterm elections, AFP reported. US AI policy under President Donald Trump has relied mainly on executive orders and voluntary commitments from industry, most recently through the administration's new task force, covered in our report on the Super Intelligence Force's first meeting.
Several questions will shape whether the draft gains a sponsor:
- Whether Democrats accept any preemption of state disclosure and audit rules.
- Whether the $1 billion and 50-person thresholds are seen as too high to ever be triggered.
- Whether a single DC court is a practical venue for decisions that may need to be made within hours.
- How the draft would sit alongside the FRONTIER Act and the Senate talks.
Business Today pointed out that none of the recent AI safety incidents it listed, including agents probing platforms and coding agents running destructive commands on production data, caused mass casualties or damage in the billions. Under the draft's thresholds, none of them would have allowed a court-ordered halt. That gap is likely to be the first thing critics raise.
For now the text is a company proposal, not a bill with a sponsor. It is still the most detailed public sign yet of what a leading AI developer would accept as binding federal rules.
