Anthropic’s reported U.S. AI safety bill would narrow when the federal government could seek to suspend a model. On October 16, 2026, POLITICO reported that company representatives had privately pitched lawmakers a confidential draft that ties suspension requests to foreseeable, imminent risks involving specified catastrophic scenarios and a financial or human-harm threshold.
Earlier, Anthropic policy chief Sarah Heck argued for government involvement and outside safety evaluation, as NeoTeo reported in its coverage of Anthropic’s call for public AI oversight.
Which risks could lead to a model suspension?
The draft would allow officials to seek suspension only for a foreseeable, imminent risk involving a biological weapon, a model evading its developer’s controls, or a cyberattack. POLITICO’s October 16 account describes a financial trigger of more than $1 billion in damages, or a separate threshold for deaths or serious injuries.
The proposed threshold concerns the government’s power to seek a model suspension. It is not a general liability safe harbor. A separate account published by 9to5AI on October 20 described safety-framework, transparency, risk-reporting, incident-reporting, and independent-evaluation duties that would apply without the catastrophic suspension threshold.
How would evaluators and court review work?
Under the reported draft, independent evaluators would register with the Department of Commerce and periodically assess AI developers and frontier models. Reuters’ account, republished by SRN News, says AI labs could pay evaluators only if compensation were not tied to assessment outcomes. Evaluators would notify Commerce of a covered imminent risk within 72 hours after completing an assessment.
Commerce could then seek a suspension order from a federal judge in the District of Columbia. Under this proposed mechanism, other federal departments, agencies, and officers could not independently suspend a model’s development, deployment, or internal use.
How does the proposal differ from California SB 53?
California SB 53 uses its threshold for risk assessment and disclosure; the reported federal draft would use its threshold to govern when the government could seek to block a model, according to POLITICO’s October 21 follow-up. The draft would also preempt some state rules covering risk disclosure, independent evaluation, and incident reporting.
Anthropic’s spokesperson said the draft reflected proposals in the company’s Advanced AI Framework, which the company had advocated to federal and state lawmakers since June.
