- Anthropic has drafted an AI safety bill for Congress that could block a model’s release if it poses a “foreseeable, imminent risk” of killing 50 or more people or causing over $1 billion in property damage.
- The draft, reported by Politico, ties qualifying risk to a model creating a biological weapon, evading developer controls, or causing a cyberattack.
- Independent evaluators registered with the U.S. Commerce Department would have 72 hours to flag an imminent risk, after which Commerce would need a Washington, D.C. judge to approve a halt.
- The bill would bar any federal agency from unilaterally blocking a model’s release, a provision distinct from how the AI industry typically frames opposition to binding rules.
What Happened
Politico reported this week that Anthropic has been shopping a draft AI safety bill to Congress, outlining scenarios in which a model’s release could be legally halted. A company spokesperson confirmed the draft, telling Politico it builds on Anthropic‘s Advanced AI Framework, which the company has pitched to federal and state lawmakers since June 2026. TechCrunch first detailed the proposal on October 20, 2026.
Why It Matters
The bill’s thresholds, 50 potential deaths or $1 billion in property damage, mirror language in California’s Senate Bill 53, a state law already in effect. Anthropic has previously pushed its own governance proposals, including the public benefit fund plan it floated in September, positioning itself as an industry voice favoring some binding oversight.
Technical Details
Under the draft, evaluators would need Commerce Department registration and could be paid by the AI labs they assess, provided compensation isn’t contingent on evaluation outcomes. Qualifying catastrophic risk is limited to three triggers: bioweapon creation, a model evading its developer’s controls, or a cyberattack. Once flagged, Commerce must petition a D.C. judge rather than act on its own authority.
Who’s Affected
The proposal would primarily affect frontier AI labs like Anthropic, OpenAI, and Google DeepMind, plus the federal agencies that would rely on third-party evaluators rather than their own judgment. Nvidia CEO Jensen Huang has separately argued the industry needs no new laws, highlighting a split in how major AI players want Washington to act.
What’s Next
Anthropic did not respond to TechCrunch’s request for comment beyond the spokesperson’s statement to Politico. The bill remains a draft with no confirmed congressional sponsor, and its evaluator-payment structure, in which labs pay the people who assess them, has not been addressed by Anthropic publicly.
