Anthropic Pitches Draft AI Bill With Narrow Court Pause

Anthropic representatives pitched US lawmakers on draft legislation that would let Washington seek a court-ordered pause on frontier AI only for narrowly defined, imminent catastrophic risks, Reuters reported. The proposal is not law.

A high threshold for intervention

Reuters reported on Oct. 16 that Anthropic representatives had pitched lawmakers on the draft, citing a Politico report based on three people familiar with the effort. The reported trigger would be a foreseeable and imminent risk of more than $1 billion in damage, or of killing or seriously injuring more than 50 people. The risk would have to arise from a biological weapon, a model evading its developer’s controls or a cyberattack.

Those thresholds concern government action to block a model’s release or use; they do not, on the available reporting, define every proposed testing or disclosure duty. The reported risk categories are:

• Creating or releasing a biological weapon.
• A model evading its developer’s controls.
• A cyberattack attributable to the model.

Under the reported draft, frontier AI developers would undergo periodic assessment by independent evaluators registered with the Commerce Department. Companies could pay evaluators, but not according to the outcome of the review. An evaluator finding an imminent risk in one of the specified categories would have to notify Commerce within 72 hours of completing its assessment.

The enforcement chain is also specific. Commerce could ask a federal judge in the District of Columbia to suspend a model. The reported language would bar other federal agencies or officials from independently suspending development, deployment or internal use. That would make judicial approval central to an emergency intervention, rather than giving one regulator a unilateral stop button.

Anthropic links the proposal to its public framework

An Anthropic spokesperson told Reuters the draft represented proposals in the company’s Advanced AI Framework, released earlier in 2026. The framework calls for frontier developers to test models, publish findings, obtain independent evaluations and maintain security programmes. It identifies biological misuse, cyber capabilities, loss of control and AI systems accelerating AI research as areas of concern.

The public framework sets coverage criteria based on both model training compute and company scale: more than 10^25 floating-point operations, plus AI-related revenue above $500 million or AI research-and-development spending above $1 billion. The reported bill’s catastrophic-risk threshold should not be confused with those coverage criteria. One describes which companies or systems would face duties; the other describes a condition for a government seeking to halt a model.

Anthropic’s stated position has also been that federal authority should not automatically erase state rules. Its June framework says Congress should pre-empt state law only if it passes a federal regime at least as strong as the company’s proposal, and that any pre-emption should be limited to the safety functions covered by federal law. The Straits Times reported that the newly circulated draft would override certain state laws, leaving a consequential question about how that reported language fits the company’s earlier position.

Congress already has a competing proposal

The draft arrives alongside H.R. 9925, the FRONTIER Act, introduced in the House on July 23 by Representative Jay Obernolte with bipartisan co-sponsors. The bill is publicly available through the Government Publishing Office and was referred to the House committees on Energy and Commerce and on Science, Space and Technology. It has not been enacted.

H.R. 9925 also uses a catastrophic-risk definition involving harm to more than 50 people or property losses above $1 billion, but it sets out a broader statutory framework for risk mitigation, independent verification, incident reporting and federal oversight. The House bill defines a frontier model using a training-compute threshold above 10^26 operations. The proposed Anthropic text, by contrast, is known through news accounts rather than a bill publicly filed with Congress, so a clause-by-clause comparison is not yet possible.

That distinction is important: matching numbers do not make the two approaches identical. The scope of covered developers, types of reportable risks, agency roles, evaluators’ independence and judicial safeguards determine how much discretion government or companies would have in practice.

Federal policy is still unsettled

The current US approach combines executive action with congressional proposals. A June 2 White House order created a voluntary process for selected frontier-model developers to provide the government with early access to some systems for cybersecurity review. The order expressly says it does not establish mandatory government licensing or pre-clearance for new models.

Anthropic’s draft would move the debate toward statutory obligations, while placing a high bar on the most intrusive step: stopping a model’s use or release. Charlie Bullock of the Institute for Law & AI, quoted by the Straits Times, described the proposal as reasonable but weaker than the FRONTIER Act. That assessment illustrates the policy trade-off: narrow intervention can guard against broad government control, but may leave less room to act before a danger meets the defined threshold.

Reuters said the Commerce Department had not immediately responded to its request for comment. The next policy debate is whether lawmakers will take up the proposal and how they balance model safety, state authority, innovation and judicial review.