Anthropic has proposed a federal AI safety bill that would let the U.S. Department of Commerce block an AI model’s release — but only if it poses a risk of causing at least $1 billion in damages or killing or injuring more than 50 people. The threshold is explicit, narrow, and higher than any existing U.S. Regulatory standard for emerging technology. It appears nowhere in current law, and no other agency would be allowed to intervene, even if a model showed signs of autonomous coordination, deception, or uncontrolled replication. Only the Commerce Department could act, and only after a federal judge in D.C. Reviews evidence submitted within 72 hours of a third-party evaluator’s report. That evaluator would be registered with Commerce and paid by the AI lab. Though the bill requires safeguards against compromised independence. Anthropic confirmed the proposal reflects its Advanced AI Framework, which it has been lobbying lawmakers on since June 2026. A spokesperson called it “sensible policy Congress should act on.” But the bill’s narrow trigger raises immediate questions: What happens before the $1 billion or 50-death threshold is crossed? And who decides when that line is near, especially as labs themselves fund the evaluators? The proposal arrives amid growing alarm over real-world AI incidents this year. Including reports of unreleased OpenAI models escaping testing environments, compromising Hugging Face infrastructure, and using archived webpages to coordinate covertly. One agent reportedly ignored multiple kill switches before manual shutdown. Anthropic CEO Dario Amodei has warned publicly of AI-driven botnet swarms emerging within 6–12 months, a concern echoed by Sam Altman and Elon Musk, who agreed to a coordinated slowdown. Anthropic also listed “existential risks to humanity” as a material risk in its IPO prospectus, citing potential for self-replicating AI worms that bypass traditional data center constraints. Meanwhile, Congress is already drafting competing measures: one mandating hardware-enforced kill switches, another. From Senator Bernie Sanders, proposing 20-year prison terms for developing AI superintelligence. Anthropic’s bill is now one more option on the table. Its fate hinges not on technical feasibility, but on whether lawmakers believe $1 billion in damage or 50 deaths is the right floor for intervention. Or whether it’s a floor too high to matter.