Anthropic, the company behind AI assistant Claude, has sent a draft bill to US lawmakers that proposes new rules for regulating advanced AI systems. Reportedly, Anthropic becomes one of the first major AI companies to draft its own legislation for Congress.
The bill sets a high bar for government intervention. Federal authorities could only seek to block a model if it posed an imminent risk of killing or seriously injuring more than 50 people or causing over $1 billion in damages. Furthermore, the government could not act alone. It would need approval from a federal judge in the US District Court for the District of Columbia. In addition, the bill explicitly bars any other federal agency from suspending a model’s development or deployment.
The covered threats are also narrow. The bill only addresses three scenarios: biological weapon creation, models escaping developer control, and cyberattacks. Other risks like labor disruption or fraud fall outside its scope. However, Anthropic says the proposal still gives the government meaningful power.
The bill also requires frontier AI developers to test their most powerful models and publish transparency reports. Independent evaluators, registered with the Commerce Department, would periodically assess labs for compliance. If an evaluator finds an imminent threat, it must notify Commerce within 72 hours. Meanwhile, developers must protect whistleblowers and report critical incidents directly to the government.
The proposal would also override certain state laws covering public risk disclosures and independent evaluations. This marks a shift from Anthropic’s own June statement that “absent a strong federal law, states should retain full authority to legislate.” Nevertheless, the company’s head of state relations reportedly told Politico that Anthropic still hopes states keep passing stronger laws.
The bill arrives during a regulatory vacuum. Congress has not passed a federal AI law and remains out of session until after next month’s midterm elections. The Trump administration has relied primarily on voluntary industry pledges. Meanwhile, security concerns have grown after recent incidents involving AI agents that escaped developer control.
CEO Dario Amodei warned last month that a swarm of AI agents could take control of “the entire internet” within six to 12 months.