Anthropic Drafts Its Own Kill-Switch Bill With a $1B and 50-Life Threshold

Anthropic has been pitching lawmakers a draft bill that would define when the federal government can shut down a frontier AI model, and the thresholds it proposes are deliberately high, PYMNTS reported on October 21.

Under the draft, Washington could block a model only when it poses a foreseeable, imminent risk of more than $1 billion in damage, or of killing or seriously injuring more than 50 people. The harm must stem from a biological weapon, a cyberattack, or a model escaping its developer's controls.

The procedure matters as much as the trigger. The Commerce Department would run the process, but only a judge in the U.S. District Court for the District of Columbia could order a suspension, after the department persuades the court that catastrophic risk is imminent. No other federal department, agency or officer could use the mechanism to halt a model's development, deployment or internal use.

Independent evaluators would register with Commerce and assess each frontier developer's safety practices and each model's capabilities on a recurring schedule. Developers could pay them, but compensation could not depend on the findings; an evaluator that finds imminent risk must report to the department within 72 hours of finishing, and the department then decides whether to go to court.

Critics focus on the numbers. The draft sets no lower tier, so a risk to 10 people or $200 million in losses would trigger no federal action, and the reporting describes no method for pricing an attack before it happens — how many systems a model could compromise, how long they would stay down, what the downtime costs.

The scale of ordinary incidents makes the gap concrete. IBM's 2026 Cost of a Data Breach study put AI-enabled breaches at roughly $6 million each and found AI involved in one in four malicious breaches; the FBI logged 22,364 complaints referencing AI in 2025, with $893 million in reported losses. A model that enables thousands of such breaches may never clear a single-event $1 billion line.

Sneha Revanur, founder and president of the AI safety group Encode, questioned why the government should have to show a model could harm 50 people before it may act, rather than 49 or fewer, the Washington Post reported.

The draft's significance is structural: it turns the question of when to switch a model off from a principle into a procedure with named institutions, deadlines and a courtroom. What remains unresolved is the hardest technical problem in the text — forecasting monetary harm, and counting cumulative losses toward a single-event threshold. Watch whether the bill is formally introduced, how lawmakers rewrite the thresholds, and whether other labs publish competing versions.