Anthropic was reported on October 20, 2026 to be pitching a draft AI safety bill to the U.S. Congress: when a model is assessed to pose a foreseeable, imminent risk of killing more than 50 people or causing over $1 billion in property damage, its release could be blocked. Politico first reported the draft, and an Anthropic spokesperson confirmed it to Politico, saying it puts into law the Advanced AI Framework proposals the company has advocated to federal and state lawmakers since June.
The line is drawn at three kinds of disaster
The draft limits blockable risk to three sources: creating a biological weapon with the model's help, the model evading its developer's controls, and the model being used in a cyberattack. The 50-person and $1 billion thresholds are not new; they closely match how California's enacted SB 53 defines catastrophic risk, and a threshold that first appeared in state law has now been carried into a federal draft. For readers, the debate has shifted from whether a line should exist to where the number sits; once written into law, every frontier model would have to pass this assessment before release.
Independent evaluators judge, a judge decides
Under the draft, risk is judged not by the lab itself but by independent evaluators registered with the U.S. Commerce Department. Evaluators may be paid by the labs, as long as payment is not contingent on the results. If an evaluator finds an imminent catastrophic risk, they have 72 hours to notify the Commerce Department, which must then ask a D.C. judge to halt the release. The draft expressly forbids any federal agency from halting distribution on its own. The design deliberately cages administrative discretion: the power to stop a model ends with a court. We previously covered the FTC's investigative demands to OpenAI and Anthropic; regulatory pressure keeps building, and this time a company has handed over the rules it says it can live with.
A lab writing its own rules is not simple
The industry is not aligned: NVIDIA CEO Jensen Huang recently argued publicly that AI needs no new laws, while Anthropic has chosen to submit a text with a mandatory blocking mechanism, going further than most peers' voluntary pledges. The weakest point remains evaluator independence: pay that is not tied to outcomes does not by itself remove an evaluator's concern about being rehired. Whether Congress takes the draft up, and whether the 50-person and $1 billion thresholds survive negotiation, is the variable to watch.
