Anthropic's leaked bill would make it harder for Washington to shut down frontier models

Summary

Anthropic's leaked draft bill limits federal intervention to only three narrow disaster scenarios (biological weapons, model escape, and cyberattacks). It proposes a threshold of $1 billion in damages or more than 50 fatalities for such intervention. Additionally, a federal judge's order in Washington, D.C. is required before any suspension action can be taken.

Why it matters

This is a rare case of a major AI company attempting to write the legal rules for its own industry. The draft sets extremely high thresholds for state action while simultaneously preempting potentially stronger state-level safety requirements. The timing is also notable, coming weeks after reports that U.S. export controls sidelined two Anthropic models. This suggests the draft may also function as a defensive shield against unilateral executive actions. How Congress resolves this issue will shape the foundation of federal AI oversight in the coming years.

Highlights

• Federal intervention is limited to three scenarios with a threshold of $1 billion in damages or 50+ fatalities. All other harms fall outside this framework.
• The Department of Commerce cannot independently halt a model. A federal judge's order in Washington, D.C. is required for this action. Other agencies are also barred from suspending development or deployment processes.
• The draft preempts state-level requirements for public risk disclosures, independent evaluations, and critical incident reporting.

Critical take

The $1 billion damage threshold, the three-item threat list, and the judicial approval requirement make meaningful state intervention virtually impossible. The provision allowing independent auditors to be paid by the companies they audit creates a conflict-of-interest structure. The note that 'compensation should not be tied to findings' does little to neutralize this structure.