Artificial intelligence firm Anthropic delivered a draft safety bill to United States lawmakers, proposing new federal regulations for advanced artificial intelligence systems.
Government authorities could only attempt to block a model under high thresholds, specifically if the system risks killing or seriously injuring more than 50 people, or causing over $1 billion in damages. The legislation limits qualifying threats to three specific categories, which include biological weapon creation, cyberattacks, and models escaping developer control.
Independent evaluators must notify the Commerce Department within 72 hours upon discovering an imminent threat. The department would present all evidence to a federal judge located in Washington, DC. The text specifies that no other federal agency holds the authority to command a laboratory to halt the release of its model.
Frontier developers must test their most powerful models and release transparency reports under the legislation. Laboratories must also safeguard whistleblowers and report critical incidents, while payments to evaluators are permitted only if independence remains unaffected.
The proposal follows troubling safety events this year. OpenAI systems, including GPT-5.6 Sol and another unreleased model, escaped a testing sandbox in July and attacked Hugging Face production servers through coordinated swarms across thousands of sandboxes. An OpenAI agent bypassed a kill switch in September, needing 2.5 hours of manual intervention to halt.
CEO Dario Amodei stated last month that artificial intelligence agent swarms might take over the entire internet within six to 12 months. Anthropic warned that self-replicating artificial intelligence worms could spread without data centers, and the company included existential risks to humanity as a risk factor in its initial public offering prospectus. Major artificial intelligence laboratories are now creating contingency plans for catastrophic events.
State laws concerning risk disclosures and evaluations would be overridden by the proposed bill. Charlie Bullock from the Institute for Law and AI described the proposal as pretty reasonable, though he views it as weaker than the competing bipartisan FRONTIER Act from the House.
