Anthropic, the company behind the Claude artificial intelligence (AI) assistant, has proposed a federal AI safety bill that would allow the US government to block the release or use of advanced AI models considered capable of causing catastrophic harm.
The company has circulated a draft of the proposed legislation among US lawmakers as Congress continues to debate how to regulate increasingly powerful AI systems. The proposal seeks to establish mandatory safety requirements for developers while setting specific conditions under which federal authorities could intervene.
The draft would allow the government to seek restrictions on an AI model if it presents a foreseeable and imminent risk of causing more than $1 billion in damage or killing or seriously injuring more than 50 people. The proposal has not yet been formally introduced in Congress.
Under the proposed framework, the US Department of Commerce would be responsible for seeking a suspension of a model that meets the specified risk thresholds. A federal judge would then determine whether the evidence justifies blocking its release or continued use.
The draft identifies three categories of threats that could trigger government intervention: the creation or release of biological weapons, an AI model escaping its developer’s controls, and cyberattacks.
The proposal would also restrict other federal departments and agencies from independently suspending the development, deployment or internal use of covered AI models under this framework.
By requiring judicial approval, Anthropic’s proposal would establish a legal process for government intervention while limiting the circumstances in which authorities could stop an AI system from being deployed.
AI Developers Would Face Mandatory Safety Requirements
Beyond government intervention, the proposed legislation would introduce additional obligations for developers of the most advanced AI models, commonly referred to as frontier AI systems.
Companies would be required to test their models, disclose safety findings to the public, submit systems for independent evaluations and maintain security programmes designed to reduce risks associated with powerful AI technology.
Independent evaluators registered with the Department of Commerce would periodically assess developers and their models. Under the draft, evaluators who identify an imminent threat meeting the specified criteria would have 72 hours after completing their assessment to report it to the department.
The proposal would allow AI developers to pay independent evaluators, provided compensation was not tied to the outcome of their assessments.
These requirements are intended to replace reliance on voluntary industry commitments with a formal system of testing, transparency and independent oversight.
Proposal Could Override Some State AI Laws
The draft also reportedly includes provisions that would pre-empt certain state-level AI safety requirements, potentially limiting the ability of individual states to establish separate rules for covered AI developers.
This aspect could affect states that have introduced or considered legislation requiring advanced AI companies to disclose risks, conduct safety assessments or implement safeguards.
The proposal comes amid ongoing disagreements in the US over whether AI regulation should primarily be handled at the federal level or through a combination of federal and state laws.
Although a national framework could establish consistent requirements across the country, limiting state authority could also reduce the scope for states to introduce additional protections.
The draft comes as US lawmakers consider different approaches to regulating advanced AI systems, including mandatory safety testing, cybersecurity requirements and independent oversight.
Anthropic has argued that the federal government needs the authority to prevent dangerous AI deployments. However, its proposed thresholds would limit intervention to narrowly defined scenarios involving potentially severe harm.
The framework therefore combines stronger safety obligations for AI developers with restrictions on when the government could block a model.
As of October 19, 2026, the proposal remains a draft circulated among lawmakers rather than an enacted law. It would need to progress through Congress before becoming legally binding.
The proposal adds to the debate over how governments can manage the risks associated with increasingly capable AI systems while establishing clear legal boundaries for developers and regulators.
Read also: Tesla unveils flying Roadster with SpaceX thruster technology