Anthropic Pitches Draft Federal AI Safety Bill to Congress

Introduction

On October 16, 2026, Politico reported that representatives of Anthropic have quietly pitched Capitol Hill lawmakers on a draft bill to regulate frontier AI safety and security. According to Politico, which obtained the confidential draft, the draft had not been previously reported; the outlet cites three people with knowledge of the efforts who spoke on condition of anonymity. TechCrunch followed up on October 20 with a summary of more of Politico's reporting.

The draft itself is not public. Every detail below about the bill's contents is as reported by Politico (or by TechCrunch summarizing Politico), not Anthropic's published text. What is public is the document Anthropic says the bill is based on: the 19-page Advanced AI Framework, published in June 2026. An Anthropic spokesperson told Politico: "This bill represents the proposals outlined in Anthropic's Advanced AI Framework, which we have been advocating for with federal and state lawmakers since June. These are sensible policies we believe Congress should act on." This review sets the reported draft beside that framework and flags where the two can and cannot be compared.

Feature Overview

What Politico reports about the draft

A narrow release-blocking trigger. According to Politico, the bill would let the government block a model's release or use only if it poses a foreseeable, imminent risk of more than $1 billion in damages, or of killing or seriously injuring more than 50 people. The risk must stem from the creation or release of a biological weapon, the model evading its developer's controls, or a cyberattack. Politico says the language mirrors California's AI laws, including SB 53, which it reports was signed into law last year.

Registered independent evaluators. TechCrunch, citing Politico, reports that the draft relies on "independent evaluators" registered with the U.S. Commerce Department. They could be paid by the AI labs, so long as the "payment is not contingent on the evaluations' results."

A 72-hour clock and a judge. Per TechCrunch's summary of Politico, an evaluator who determines that an imminent catastrophic risk exists would have 72 hours to contact the Commerce Department. Commerce would then have to ask a D.C. judge to halt the model's release, and the legislation would expressly forbid any federal agency from having discretion to halt distribution on its own.

State-law preemption. Politico's summary line says the draft would limit the circumstances in which the government could block a model's release and "preempt some state AI safety laws." We could read only the opening of Politico's article, so the scope of that provision is not covered here.

What Anthropic's published framework says

The framework describes its enforcement section as "a range of options for policymakers" that can be "mixed and matched," and suggests policymakers could begin with a lighter-touch model.

Draft Versus Framework: Our Analysis

The points below are our reading, not statements by Politico, TechCrunch or Anthropic.

Where they line up. The framework lists court enforcement (an agency pursues remedies by filing a lawsuit) and cabined discretion (no action based on the agency's own risk assessment) as safeguards against overreach. The reported structure, in which Commerce must ask a D.C. judge and no agency may halt distribution on its own, resembles both. Both documents also anchor the catastrophic-risk concept to SB 53.

Where they differ on the surface. The reported block trigger names three risk types, while the framework prioritizes four, adding automated research and development. Whether the draft addresses that fourth category elsewhere is not in the reporting we read. The 72-hour evaluator notification also does not appear in the framework text, which uses a 15-day window for developer incident reports, a different obligation.

Evaluator payment. The framework says an independent evaluator should have no financial interest in the developer, and it discusses government or pooled funding so evaluators stay financially independent of any given developer. The reported draft permits lab payments that are not contingent on results. Whether a fee counts as a financial interest depends on definitions in the draft text. TechCrunch notes it is not clear how the system would assure that payments could not influence an evaluation.

Preemption. The framework says Congress should not preempt state law unless it enacts a rigorous federal regime that meets or exceeds the framework's strongest measures, and even then only for specific functions such as catastrophic-risk testing, evaluator licensing, or closely related reporting. It adds that compliance should not itself confer immunity or a safe harbor. Politico's one-line summary of the draft's preemption cannot be tested against those conditions. Whether a bill whose blocking authority Politico describes as limited meets the framework's "meets or exceeds" bar is a question only the full text could answer.

Usability Analysis

This is a draft in circulation, not a law. Neither the portion of Politico's article we could read nor TechCrunch's follow-up names a sponsor or reports that a bill has been introduced, so there is nothing for developers to comply with yet.

For readers tracking the policy, the practical use of this story is as a checklist. If the draft follows the framework, coverage would turn on the 10^25 FLOP and revenue or spending tests, which would reach only very large developers. That is conditional: the reporting does not say the draft adopts those thresholds. TechCrunch also offers its own assessment that the draft's scenarios are somewhat removed from recent industry concern about agentic hacking and rogue AI incidents, and that Anthropic's proposal appears to go slightly further than other industry proposals, noting that Nvidia's CEO has argued the industry does not need new laws.

Pros and Cons

On the positive side, a public framework exists to read the draft against, and an Anthropic spokesperson confirmed the link in a quoted statement. The reported process routes any halt through a judge, a safeguard the framework itself names, and mirrors language from an existing state law, per Politico, which gives lawmakers a familiar reference point.

On the limitations side, the draft is confidential, so most details rest on press reporting rather than a primary document. The evaluator payment rule leaves open how independence is protected. The preemption scope is not detailed in the reporting available to us, and the reported trigger covers three risk types against the framework's four.

Outlook

Several things would make the picture clearer: publication or leak of the full text, a named sponsor, and detail on the preemption provision and the evaluator definitions. Until then, the reported draft and the framework are best read together, with the framework as the stated basis and the reporting as a partial view of how it was turned into statutory language.

Conclusion

Anthropic's reported draft is a notable attempt by a frontier lab to put its own policy proposals into bill form, according to Politico. The framework behind it is public and detailed; the bill is not. Readers interested in AI policy, evaluator independence or federal-state preemption should follow how the open questions above are resolved. Our rating reflects solid sourcing for the framework and a partial view of the draft.

Editor's Verdict

Anthropic Pitches Draft Federal AI Safety Bill to Congress is a story worth following, with the caveat that the bill itself remains out of public view.

The strongest case for paying attention: a public framework exists to read the draft against, and an Anthropic spokesperson confirmed the link in a quoted statement. That gives readers a published yardstick rather than only secondhand descriptions. Reinforcing that, the reported process routes any halt through a judge rather than agency discretion, a safeguard the framework itself names — practical value rather than just headline appeal. The broader signal worth registering is straightforward: the draft is confidential, so the bill's contents come from Politico's reporting, and the public framework is the primary document available for comparison. On the other side of the ledger, one constraint is real rather than a marketing footnote: the draft text is confidential, so most details rest on press reporting rather than a primary document. It should factor into any serious decision. Layered on top of that, the reported evaluator payment rule leaves open how independence is protected, a gap TechCrunch also flags — which narrows the set of teams for whom this is an obvious yes.

For policy watchers, compliance teams at large AI developers, and readers following the federal-versus-state debate, the practical move is to read the June framework now and compare it against the full bill text if and when it becomes public. For everyone else, nothing changes yet: this is a circulating draft, not legislation.