▶ Anthropic has informally submitted draft legislation on advanced AI safety regulation to members of the U.S. Congress.
▶ The draft limits the sources of risk to the creation or release of biological weapons, AI systems escaping developer control, and cyberattacks.
▶ It also includes provisions that would replace state laws on risk disclosure, external assessments, and incident reporting with federal law.
Anthropic has drafted an AI regulation bill and presented it to Congress. Politico reported on the 16th, citing three people familiar with the matter, that Anthropic representatives had informally submitted draft legislation to members of Congress to regulate the safety and security of advanced AI.
The proposal would allow the government to block the release or use of an AI model when large-scale harm is imminent, while narrowly defining the conditions for intervention and who may intervene.
An Anthropic spokesperson said the draft would give the government authority to “block or contain the dangerous deployment of the most powerful models.”
◆ Intervention requires both a specified level of harm and a qualifying cause
The draft sets numerical thresholds for government intervention: a foreseeable and imminent risk of property damage exceeding $1 billion (about 1.385 trillion won), or more than 50 deaths or serious injuries.
It also limits the potential causes of harm to three: ▲ AI being used to create or release biological weapons; ▲ a model escaping its developers’ control; and ▲ involvement in a cyberattack. The government could intervene only if one of these conditions applied and harm at the specified scale was imminent.
This summer, major AI companies disclosed incidents involving AI agents navigating websites and platforms or engaging in hacking. So far, none of these incidents has caused deaths or harm on the scale specified in the draft.
◆ Commerce Department requests action; court makes the decision
The authority to halt a model would rest with the courts, not the executive branch. If the Commerce Department determines that a risk exists and requests a halt, a judge on the U.S. District Court for the District of Columbia could order the model’s release or use to be suspended.
The draft would prohibit other departments, agencies, and government officials from halting the development, deployment, or internal use of AI models. This provision would prevent agencies such as the Federal Trade Commission (FTC) from independently stopping an AI model.
This structure intersects with other discussions underway in Congress. On September 29, Senate Commerce Committee Chairman Ted Cruz opposed a request for unanimous consent to advance the AI Risk Management and Security Act, introduced by Democratic lawmakers including Mark Warner. Cruz argued that the bill gave too much authority to the Commerce Department and a safety commission, and said he preferred a role for “an independent judge.”
The Warner bill would establish an AI safety commission within the Commerce Department and require developers to provide access to models 45 days before release. Cruz is negotiating a separate bill with Senate Majority Leader John Thune and Senator Amy Klobuchar that would focus on biological and nuclear disaster risks.
◆ External assessments and reporting within 72 hours
The draft also sets out obligations for developers. Under the proposal, developers of advanced AI would have to test their models’ safety and publish the results in transparency reports.
It would protect whistleblowers and require serious safety incidents to be reported to the Commerce Department. External assessment organizations registered with the department would regularly inspect AI labs and models.
If an assessment organization identifies an imminent risk as defined by the bill, it would have to notify the Commerce Department within 72 hours of completing its assessment. Labs could pay for the assessments, but compensation could not vary based on the findings.
The draft also includes provisions that would replace some state laws with federal law in areas including risk disclosure, external assessments, and reporting of serious incidents.
In its Advanced AI Framework, published in June, Anthropic said state laws should not be preempted unless the federal government adopts regulations at least as rigorous as the framework.
◆ Turning the June framework into legislative language
An Anthropic spokesperson told Politico that the draft reflects proposals in the company’s Advanced AI Framework, released earlier this year, and described it as “a sensible policy for Congress to act on.”
Anthropic said it had been explaining the framework to federal and state lawmakers since June. The framework proposed requiring developers that train models exceeding 10^25 FLOPs (floating-point operations) and have more than $500 million in AI revenue or $1 billion in research and development spending to publish a safety framework, issue risk reports every six months, report serious incidents within 15 days, and involve independent assessment organizations.
Similar bills are already before Congress. In July, Representatives Lori Trahan (Democrat) and Jay Obernolte (Republican) introduced the Frontier Act in the House, which would allow the government to halt the deployment of dangerous models.
Some lawmakers and AI safety groups argue that the government should play a much larger role in reviewing advanced models.
Last month, President Trump signed a voluntary agreement with executives from six companies, including Anthropic, Google, Meta, Nvidia, and OpenAI, pledging internal controls, external assessments, and board oversight. The House is scheduled to return on November 9, after the November 3 midterm elections. No lawmaker has yet agreed to sponsor Anthropic’s draft.