Anthropic Is Quietly Writing Its Own AI Safety Law, and the Bar for Using It Is Sky-High

Anthropic representatives have quietly pitched Capitol Hill lawmakers on a draft bill to regulate frontier AI safety and security, according to a Politico report published Friday, October 16, 2026, and carried by Reuters. Citing three people with knowledge of the effort, the report describes a proposal that would give the government a narrowly drawn but extraordinary power: blocking the release or use of a frontier AI model, but only when it poses a foreseeable, imminent risk of more than $1 billion in damages, or of killing or seriously injuring more than 50 people.

The draft would build an evaluation regime around that power. Independent evaluators would register with the Department of Commerce and periodically assess the lab and its frontier models. Labs could pay those evaluators, provided the payment is not tied to the evaluation's outcome. If an evaluator determines there is what the draft calls "imminent enumerated risk," it must report that finding to Commerce within 72 hours of completing its assessment. From there, Commerce could ask a judge to halt the model, but the draft explicitly prohibits any other U.S. government department, agency, or officer from suspending a model's development, deployment, or internal use. A federal judge for the District Court for the District of Columbia would make the final call.

An Anthropic spokesperson told Reuters the bill represents proposals outlined in the company's "Advanced AI Framework," published earlier this year, and said it would grant the government the "ability to block or deter dangerous deployments of the most powerful models," while requiring developers to test frontier models, be transparent about their findings, and submit to independent evaluation. As of October 16, 2026, no member of Congress has introduced the draft as legislation, and the Commerce Department did not respond to Reuters' request for comment. This article is explanatory only, not legal advice.

The threshold that does all the work

The $1 billion and 50-person thresholds are the heart of the proposal, and they reveal almost everything about what it is designed to catch. The qualifying risk must stem from one of three triggers: the creation or release of a biological weapon, a model evading its developer's controls, or a cyberattack. Everything else that AI regulation usually argues about, biased hiring tools, misinformation at scale, data scraping, labor displacement, falls outside the draft's reach by design.

That narrowness is a choice with a theory behind it. In Anthropic's framing, this is catastrophic-risk regulation, not general-purpose AI governance. The thresholds are meant to separate existential-scale scenarios from the ordinary consumer harms that existing law, from FTC deception authority to state privacy statutes, already claims to cover. The bet is that lawmakers will accept a bill that ignores most AI harms if it credibly addresses the worst ones.

Skeptics read the same thresholds differently. A billion dollars in foreseeable, imminent damage is a bar so high that almost no incident in AI's history would have cleared it. If the thresholds are never met in practice, the bill becomes a framework that looks like regulation while changing very little about how labs operate.

Anthropic is asking Congress for a power the government might never use: the authority to stop a model only when the damage is already measured in billions of dollars or dozens of lives.

A bottleneck, built on purpose

The draft's most consequential clause may be the one that limits who can act. Only the Commerce Department could request a halt, and only a federal judge could order one. Every other agency would be barred from suspending a model's development, deployment, or internal use, no matter what it found.

Read alongside the rest of the draft, this is a preemption machine dressed as due process. One agency, one judge, one court in Washington, D.C. For a lab facing a patchwork of federal enforcers, that is a feature: a single forum replaces the risk of open-ended intervention by any agency with jurisdiction. It is also the clause most likely to draw fire from regulators who see their own authority written out of the statute.

The contrast with what the FTC is doing right now is sharp. One day before the Politico report, the FTC confirmed to CBS News that it had issued civil subpoenas to OpenAI and Anthropic in a consumer protection investigation into AI safety incidents, seeking records on safety incidents, safeguards, marketing practices, and privacy policies. The FTC's theory is the adversarial one: gather evidence of deceptive safety claims and punish violations after the fact. Anthropic's draft proposes the opposite posture, a cooperative regime where the government acts only through registered evaluators, 72-hour reports, and court orders.

Regulation written by the regulated

The deepest question is structural. This is a draft written by the company it would regulate, pitched quietly to lawmakers by its own representatives. Anthropic's argument is that it understands frontier risk better than anyone else, which is why its "Advanced AI Framework" should become the template for law. Critics will note the familiar pattern: industries that write the rules of their own supervision, from finance to telecom, usually produce rules they can live with.

The evaluator design invites the same scrutiny. Labs paying their own evaluators, even with outcome-independence rules, creates a relationship the industry calls independent and everyone else calls captured. The draft tries to solve the conflict with a firewall: payment allowed, outcome-tied payment banned, plus Commerce registration. Whether that firewall holds depends on details the public has not seen, the registration standards, the evaluation methodologies, and who enforces the independence rules.

It is also worth noting what the draft does not touch. There is no transparency requirement aimed at the public, no provision for state attorneys general, no role for the courts beyond the single halt-order mechanism, and no answer for models developed abroad that cause the same harms. Meanwhile, state law is moving on an entirely separate track: Connecticut's new AI provisions, which took effect October 1, 2026 and include whistleblower protections for frontier developers and content provenance duties, build regulation use case by use case, state by state.

What to watch next

As of October 16, 2026, the draft is a pitch, not a bill. It has no sponsor, no number, and no committee referral. The timing matters: with the midterm elections weeks away, major legislation is unlikely to move before the lame-duck session, and even sympathetic lawmakers may treat the draft as a starting position rather than a final text. The real test is whether any lawmaker introduces it, and what survives first contact with committees.

A second federal track is already in motion. On October 12, Senators Mark Warner, Brian Schatz, and Andy Kim introduced the Artificial Intelligence Risk Management and Security Act, which would create an AI Safety Board inside the Commerce Department to develop evaluation standards and testing protocols, without any power to halt a model. Congress now has two competing templates: Warner's board, which studies and sets standards, and Anthropic's draft, which empowers a judge to pull the plug. Which one lawmakers reach for will say a lot about whether Washington wants to understand AI risk, or simply to stop it at the brink.