Anthropic Pitches AI Safety Bill to US Lawmakers: Key Proposals Explained

Anthropic has quietly presented US lawmakers with a draft bill outlining how the federal government should regulate the safety and security of advanced artificial intelligence systems, according to a Politico report published on October 16, 2026, and cited by Reuters. The proposal would establish a specific legal threshold for government intervention, allowing authorities to seek a halt to an AI model’s release or use only when it presents a foreseeable and imminent risk of causing more than $1 billion in damage or killing or seriously injuring more than 50 people under defined circumstances.

The proposal focuses on threats involving biological weapons, AI systems evading their developers’ controls and cyberattacks. It would also introduce independent evaluations of frontier AI models, a 72-hour reporting requirement for certain imminent risks and a more narrowly defined process for seeking judicial intervention. The draft is not an enacted law, and its provisions could change if lawmakers consider or amend it. Nevertheless, the initiative highlights an important debate in the AI industry: how to regulate potentially dangerous systems while preserving the ability to develop and deploy useful technology.

Key Takeaways

• Proposed legislation: Anthropic has reportedly presented lawmakers with a draft bill covering frontier AI safety and security.
• Financial threshold: Government intervention could be sought for a foreseeable, imminent risk exceeding $1 billion in damage.
• Human-harm threshold: The proposal also covers risks involving the death or serious injury of more than 50 people.
• Specified threats: The draft focuses on biological weapons, models evading developer controls and cyberattacks.
• Independent evaluations: Frontier AI models would undergo periodic assessments by independent evaluators registered with the Commerce Department.
• 72-hour reporting: Evaluators identifying a qualifying imminent risk would have to report it to the Commerce Department within 72 hours of completing their assessment.
• Legal oversight: A federal judge could order a model’s suspension under the proposed framework.
• Current status: The legislation remains a proposal, not a law approved by Congress.

What Is Anthropic Proposing?

Anthropic, the developer of the Claude family of AI models, has been involved in discussions about how governments should manage the risks associated with increasingly capable AI systems.

Its reported draft bill seeks to create a defined process for dealing with frontier AI models—advanced systems whose capabilities may introduce risks beyond those associated with ordinary software.

The central idea is to set a relatively high threshold for government intervention. Rather than allowing officials to stop a model simply because it raises general concerns, the proposed framework would require a foreseeable and imminent threat that meets specified criteria.

Under the draft described by Politico, the risk would need to involve either more than $1 billion in damage or the death or serious injury of more than 50 people. The threat would also have to arise from one of the specified categories of danger.

The financial and human-harm thresholds are important because they attempt to define when a risk becomes serious enough to justify extraordinary regulatory action. They are not, however, general limits on all harm that AI systems could cause, nor do they establish that every incident below the thresholds would be legally irrelevant.

The proposal would instead create a framework for a particular type of intervention involving advanced AI systems and specified imminent threats.

The Three Risks That Could Trigger Government Action

The reported bill identifies three broad categories of risk that could qualify for intervention.

1. Creation or release of biological weapons

The first category involves the creation or release of a biological weapon. Advanced AI systems can provide information that helps people understand complex scientific processes, which creates potential benefits for legitimate research as well as risks if capabilities are misused.

The proposal would focus on circumstances in which an AI model creates a foreseeable and imminent threat connected to biological weapons and meets the bill’s other conditions.

This is narrower than treating every AI application involving biology as dangerous. Scientific research, medical development and other legitimate applications can involve biological information without constituting a weapons threat.

2. An AI model evading its developer’s controls

The second category concerns an AI model evading the controls imposed by its developer.

AI safety researchers study whether advanced systems could act in ways that conflict with their intended restrictions, particularly when given access to tools, software environments or other capabilities.

The proposed legislation would treat a qualifying, imminent control-evasion threat as one potential basis for government intervention. The specific legal determination would still depend on the evidence and requirements established by the final legislation, if enacted.

The provision reflects a broader concern about whether developers can reliably constrain the behaviour of increasingly capable systems after deployment.

3. Cyberattacks

The third category involves cyberattacks. AI systems can assist with legitimate cybersecurity work, including identifying vulnerabilities, analysing software and helping defenders respond to incidents. The same general capabilities can also be misused.

The draft would focus on an imminent cyberattack-related threat that satisfies its specified harm threshold and other conditions.

This distinction matters because the proposal is not described as a blanket prohibition on AI-assisted cybersecurity tools. Instead, it seeks to establish circumstances in which the government could seek to prevent the release or use of a model presenting a qualifying danger.

How Independent AI Evaluations Would Work

Another major element of the proposal is a system of independent evaluations for AI developers and their frontier models.

According to the reported draft, evaluators would register with the US Department of Commerce and periodically assess AI laboratories and their advanced models. These evaluations would be intended to identify risks before they result in serious harm.

The arrangement would also allow AI companies to pay the evaluators, provided the payment was not tied to the outcome of an assessment.

That provision raises an important question about independence. AI companies have specialised technical knowledge and the resources needed to support detailed assessments, but evaluators must also be able to report problems without pressure to produce favourable findings.

The draft’s proposed payment condition is intended to separate compensation from the result. How that principle would be enforced, what qualifications evaluators would need and how conflicts of interest would be handled would depend on the eventual legal and administrative rules.

Independent evaluations could give regulators more structured information about advanced AI systems. Their effectiveness would nevertheless depend on access to relevant evidence, testing methods, evaluator competence and the ability to identify risks that may emerge only under particular conditions.

Why the 72-Hour Reporting Requirement Matters

The proposed legislation would require independent evaluators to report a qualifying imminent risk to the Commerce Department within 72 hours of completing their assessment.

The deadline would create a formal reporting obligation rather than leaving the timing entirely to an evaluator’s discretion.

This matters because the value of an AI safety assessment depends partly on whether important findings reach the appropriate authorities in time for action. A detailed evaluation that identifies a serious and imminent danger may have limited practical value if the findings are not communicated promptly.

However, the 72-hour period described in the report begins after the assessment is completed. It should not be interpreted as a universal requirement to discover, investigate and resolve every possible AI risk within three days.

The proposal also does not mean that every reported risk would automatically lead to a model being suspended. The reported framework places additional conditions on government intervention and judicial action.

Who Would Have the Power to Stop an AI Model?

The proposed bill would limit the circumstances in which the federal government could seek to stop a model’s release or use.

According to the reported draft, the Commerce Department could request judicial intervention after determining that a qualifying risk exists. A federal judge in the US District Court for the District of Columbia could then order a suspension under the proposed framework.

The draft reportedly restricts other federal departments, agencies and officials from independently suspending a model’s development, deployment or internal use under this particular process.

This would make judicial review a central part of the proposed intervention mechanism. Instead of allowing a broad range of government bodies to halt a model independently under the bill, the framework would direct the process through a designated department and court.

The practical effect would depend on the final wording, the evidence required to establish risk and how courts interpreted the legislation. Because the bill has not been enacted, the proposed allocation of authority should not be confused with a description of all existing federal powers over AI or other areas of law.

The Debate Over Federal and State AI Regulation

The reported proposal also raises questions about the relationship between federal AI rules and state-level legislation.

According to reporting on the draft, it could pre-empt some state AI safety laws. Pre-emption generally means that federal law limits or displaces certain state requirements when the legal conditions for doing so are met.

Supporters of a national framework may argue that companies developing advanced AI systems need consistent rules across the country. A single federal standard could reduce conflicting obligations and make compliance more predictable for businesses operating in multiple states.

Critics of broad pre-emption may argue that states need room to respond to local risks, address gaps in federal rules or adopt stronger protections when circumstances warrant them.

The balance is especially important in AI because the technology is used across many industries and jurisdictions, from healthcare and education to cybersecurity and financial services.

The precise scope of any pre-emption provision would depend on the final legislative text. It would therefore be premature to conclude that the proposal would eliminate all state-level AI regulation.

Why Anthropic Is Taking This Approach

Anthropic has publicly emphasised AI safety as part of its approach to developing advanced models. Its reported proposal appears to translate some of that focus into a specific regulatory framework.

The draft also reflects a challenge facing AI developers: they must address genuine safety concerns while operating in a competitive market where the timing and availability of new capabilities matter commercially.

A defined intervention threshold could give companies more clarity about when the government might seek to stop a release. Independent evaluations could also provide a formal way to assess certain risks.

At the same time, a company-authored proposal naturally reflects its preferred approach to regulation. Lawmakers and independent experts would need to examine whether its thresholds are appropriate, whether the evaluation system is sufficiently robust and whether the proposed limits on government authority adequately protect the public.

Anthropic’s proposal should therefore be treated as a contribution to the policy debate, not as a settled consensus on how frontier AI should be regulated.

What the Proposal Could Mean for AI Companies

If a similar framework were eventually enacted, AI developers could face clearer requirements around evaluations, risk reporting and the circumstances in which authorities could seek judicial intervention.

Companies might need formal processes for identifying qualifying risks, preserving assessment records, communicating with registered evaluators and reporting findings to regulators.

The framework could also affect how companies plan model releases. If a model raised a credible concern involving one of the specified threats, developers and evaluators would need to consider the applicable legal process and reporting obligations.

However, the proposal’s narrow intervention thresholds would not necessarily remove other compliance requirements. AI businesses could still face applicable laws involving privacy, consumer protection, cybersecurity, intellectual property and other areas.

The business implications would depend on the final legislation, any implementing regulations and how the rules interacted with existing federal and state requirements.

For smaller AI developers, the cost of formal evaluations could also become a consideration if the framework applied to them. The draft’s reported provisions and eventual implementing rules would need to clarify which developers and models qualify for the requirements.

The Bigger Picture

Anthropic’s draft bill highlights a central challenge in AI governance: regulators need ways to act when advanced systems pose serious, imminent dangers, but intervention standards must also be sufficiently clear to avoid unnecessary restrictions on research and useful applications. The proposed harm thresholds, independent evaluations and judicial process are attempts to define that balance.

The proposal is not yet law, and its eventual influence will depend on whether lawmakers take it up, how they revise its provisions and how it fits into the wider debate over federal and state authority. Its significance lies in putting concrete rules on the table for discussion—not in establishing a new legal regime immediately.

Looking Ahead

The next steps will depend on whether lawmakers formally consider Anthropic’s draft and whether other AI developers, safety researchers, civil-liberties groups and industry organisations offer competing proposals. Important questions include how independent evaluators would be accredited, what evidence would establish an imminent risk, how conflicts of interest would be managed and how courts would assess requests to suspend a model.

For the AI industry, the debate could help determine the regulatory expectations surrounding the most capable systems. For the public, the key issue will be whether the eventual framework can provide meaningful protection against severe AI-enabled threats while maintaining transparency, accountability and appropriate oversight of government intervention.

Frequently Asked Questions

1. What AI safety bill has Anthropic proposed?

Anthropic has reportedly presented US lawmakers with a draft bill to regulate frontier AI safety and security. It proposes defined thresholds for government intervention, independent evaluations and a formal risk-reporting process.

2. When could the government block an AI model under the proposal?

The reported draft would permit intervention when a model presents a foreseeable, imminent risk of more than $1 billion in damage or the death or serious injury of more than 50 people, arising from specified threats such as biological weapons, control evasion or cyberattacks.

3. What is the 72-hour rule?

Independent evaluators who identify a qualifying imminent risk would have to report it to the Commerce Department within 72 hours of completing their assessment, according to the reported proposal.

4. Would the Commerce Department be able to shut down an AI model directly?

The reported framework would allow the Commerce Department to seek a court order after determining that a qualifying risk exists. A federal judge could order suspension under the proposed process. The draft reportedly limits independent suspension powers for other federal agencies under this mechanism.

5. Is Anthropic’s AI safety bill already law?

No. It is a draft proposal reportedly presented to lawmakers. It would need to go through the relevant legislative process before becoming law.

6. Could the proposal affect state AI laws?

The draft reportedly includes provisions that could pre-empt some state AI safety laws. The precise scope and legal effect would depend on the final wording and any legislation enacted.