Anthropic Quietly Pushes AI Safety Bill That Would Limit Government Power to Block Dangerous Models

Anthropic, the company behind the Claude AI chatbot, has privately pitched lawmakers its own federal AI safety bill that would restrict when the government can block dangerous artificial intelligence models, even as it establishes new oversight requirements for the industry.

Under the proposal, federal officials could seek to suspend an advanced AI model only if it poses a foreseeable and imminent risk of causing more than $1 billion in damages or killing or seriously injuring more than 50 people. Politico first reported the legislative effort on October 16, 2026, citing three people familiar with the discussions.

The proposed thresholds would apply to three categories of threats: the creation or release of biological weapons, cyberattacks, and AI systems escaping their developers' controls. The legislation targets frontier AI models, a term generally used to describe highly advanced artificial intelligence systems capable of performing increasingly complex tasks.

An Anthropic spokesperson confirmed that the draft reflects proposals outlined in the company's Advanced AI Framework released earlier this year. The company said its approach would allow the government to intervene against dangerous deployments while requiring developers to test their most powerful systems, disclose safety findings and undergo independent evaluations.

Under the draft, outside evaluators registered with the Commerce Department would periodically assess AI developers and their frontier models for potential threats. Companies could pay those evaluators, provided their compensation was independent of the findings.

If an assessment identified an imminent threat meeting the legislation's criteria, evaluators would have 72 hours after completing their review to report the findings to the Commerce Department.

Federal officials could then seek a court order suspending the model. A judge in the U.S. District Court for the District of Columbia would decide whether the identified risk justified intervention, making judicial authorization a prerequisite for a suspension under the proposed process.

Other federal departments, agencies and officials would be barred from independently suspending the development, deployment or internal use of covered AI models through that framework. The restriction would concentrate the authority to seek such action within the Commerce Department, with federal judges responsible for deciding whether to grant it.

The draft also reportedly includes provisions that would override certain state AI safety laws, potentially limiting states' ability to impose separate testing, disclosure and safety requirements on developers. Known as federal preemption, the legal principle allows federal law to supersede state law in areas Congress has chosen to regulate. The extent of those restrictions would depend on the final legislative language.

Anthropic's effort comes as lawmakers pursue other proposals to govern the rapidly advancing technology.

On September 29, 2026, Democratic Sens. Mark Warner of Virginia, Brian Schatz of Hawaii and Andy Kim of New Jersey introduced the Artificial Intelligence Risk Management and Security Act of 2026. The measure, S. 5576, would establish a federal Artificial Intelligence Safety Board and create enforceable safety and security requirements for advanced models.

The Senate proposal would require developers to conduct pre-release safety testing, maintain cybersecurity safeguards and report significant incidents. Warner sought unanimous consent to pass the bill on September 29, but Republican Sen. Ted Cruz of Texas objected, blocking its passage through that procedure. The measure remains before the Senate Committee on Commerce, Science, and Transportation.

Anthropic has previously supported state-level AI safety legislation, including California's frontier AI transparency requirements. Its effort to establish national rules comes as lawmakers consider how federal oversight should interact with state protections and how much authority regulators should have over the industry's most powerful systems.

The company is also expanding independent scrutiny of its technology. On October 15, Anthropic said it had granted six researchers from the nonprofit Model Evaluation and Threat Research (METR) employee-like access to internal systems, company hardware, and office space. The arrangement allows researchers to examine safety practices and assess risks associated with advanced AI models.

METR can publish its findings independently, although Anthropic retains limited rights to require redactions involving sensitive information. The company has also been working with SecureBio and Accenture on external assessments, reinforcing its broader push for third-party evaluations as a central component of AI oversight.

The privately circulated legislative draft has yet to be formally introduced in Congress. Its proposed limits on federal intervention and state regulatory authority would face congressional consideration if lawmakers agree to sponsor the measure. Any legislation would require approval from both chambers and the president's signature before becoming law.