The draft legislation, obtained by Politico and not previously reported, would allow the government to block the release or deployment of an AI model only if it posed a foreseeable and imminent danger of causing more than $1 billion in damage, or killing or severely injuring more than 50 people.
The risk would have to arise specifically from cyberattacks, the creation of biological weapons or a model evading its developer's controls.The proposal mirrors language in California laws, including SB 53, the report added.
It comes as the industry deals with safety lapses, including routine disclosures of AI agents probing platforms, autonomous test models escaping sandbox environments to probe external networks, malicious actors using AI tools for autonomous cyber-espionage, and AI coding agents executing unauthorised destructive commands on production data.
None of these incidents caused catastrophic casualties or destruction worth billions of dollars, but they have intensified the debate over where regulatory limits should be set.
To enforce compliance, the draft would establish a system under which independent evaluators registered with the Commerce Department periodically assess frontier AI laboratories and their models. AI companies could pay the auditors directly, provided their compensation was not linked to the results of the assessments.
If an evaluator identified an "imminent enumerated risk", the findings would have to be reported to the Commerce Department within 72 hours.
Commerce officials would not be allowed to shut down a model unilaterally even if they determined that it posed an imminent risk. The department would instead have to ask a federal judge in the District Court for the District of Columbia to order its suspension.
The draft would also prohibit any other federal agency, department or officer from independently halting a model's development, deployment or internal use. Beyond these judicial restrictions, the bill would override state and local rules covering public risk disclosures, third-party audits and the reporting of critical incidents, as per the Politico report.
Federal pre-emption remains a political point of dispute on Capitol Hill, with Democrats largely resisting efforts to weaken safeguards enacted by states such as California, Illinois and New York. Technology groups, meanwhile, argue that differing state regulations create an excessive compliance burden.
The proposed legislation would also require frontier AI developers to establish safety frameworks, publish transparency reports for new models, maintain whistleblower protections and report critical safety incidents directly to the Commerce Department.
Anthropic's proposal combines safety and transparency requirements with a threshold of more than $1 billion in damage or over 50 deaths or severe injuries for intervention, independent assessments and judicial approval for suspensions, while seeking to replace some state-level rules with a federal framework.