Anthropic's draft federal AI safety bill: what is reported about court-gated model blocks

Anthropic has been circulating a draft federal bill on frontier AI safety among lawmakers in Washington, according to reporting by Politico on 16 October 2026 that an Anthropic spokesperson confirmed. The draft is not public and has no known sponsor, so everything below comes from reporters' descriptions of it. The core idea: the government could block a model's release or use 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, and only through a federal judge. It matters because it is a leading AI lab writing the terms on which it could be regulated, at a moment when Congress has no federal AI law.

Key points

• Politico reported on 16 October 2026 that Anthropic quietly pitched a draft frontier AI safety bill to lawmakers; an Anthropic spokesperson confirmed that it reflects the company's Advanced AI Framework.
• Per the reporting, a model could be blocked only for a foreseeable, imminent risk of over $1 billion in damages, or of killing or seriously injuring more than 50 people, from a biological weapon, a model evading its developer's controls, or a cyberattack.
• Independent evaluators registered with the Commerce Department would assess labs and models, could be paid by the labs if payment is not tied to results, and would have 72 hours to report an imminent risk to Commerce.
• Commerce would have to ask a federal judge in the District of Columbia to halt a model, and the draft would reportedly bar other federal agencies from suspending development, deployment or internal use.
• The bill has not been introduced, and separate Senate talks involving Anthropic are reported to be on hold.

What happened

On Friday 16 October 2026, Politico reported that Anthropic representatives had been pitching a draft AI safety bill on Capitol Hill, citing three people with knowledge of the effort. Reuters and TechCrunch then summarised the report. Anthropic did not immediately respond to TechCrunch's request for comment, but a spokesperson confirmed the draft to Politico: "This bill represents the proposals outlined in Anthropic's Advanced AI Framework," which the company says it has been advocating with federal and state lawmakers since June. Another statement attributed to the spokesperson describes the bill as giving the government the "ability to block or deter dangerous deployments of the most powerful models."

We have not seen the draft itself. The descriptions below come from Politico's reporting as relayed by Reuters (via The Star) and TechCrunch, and where outlets differ we say so.

When a model could be blocked. The government could act only if a model poses a "foreseeable, imminent risk" of more than $1 billion in property damage, 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.

Who decides. Independent evaluators would register with the Commerce Department and periodically evaluate each frontier developer and its models. Labs may pay them, provided the payment is "not contingent on the evaluations' results." If an evaluator finds an "imminent enumerated risk," it must report to Commerce within 72 hours of finishing the assessment. Commerce could then ask a judge in the US District Court for the District of Columbia to halt the model. The draft explicitly bars other federal departments, agencies or officers from suspending development, deployment or internal use.

What developers must do. Anthropic says developers would have to test their most powerful models, be transparent about the results, and submit to independent evaluation.

Background

The thresholds look familiar. TechCrunch notes that the $1 billion and 50-person figures resemble those in California's Senate Bill 53, the recently passed state law that defines "catastrophic risk" in a similar way.

The proposal also builds on a theme Anthropic has pushed for weeks: independent evaluators. On 15 October, a day before the Politico story, Anthropic published "An update on independent evaluators," saying it had begun giving six employees of the nonprofit METR employee-like access to company systems, with METR free to publish its findings. Anthropic says it cannot veto METR's publications but can require redaction of predefined categories such as confidential intellectual property. The post also names SecureBio, which is reviewing the chemical and biological weapons chapter of Anthropic's August 2026 risk report, and says Accenture and Faculty are working toward a broader partnership. Anthropic argues in that post that "every frontier AI company should do the same", and compares the model to bank supervisors who place resident examiners inside the largest banks. The legislative draft is, in effect, a way to make that approach a legal requirement.

Elsewhere, positions differ. TechCrunch recalls that DeepMind chairman Demis Hassabis proposed in July a self-governing industry body modelled on FINRA, the US financial industry's self-regulator, and that Nvidia CEO Jensen Huang argued in September that the AI industry needs no new laws. The labs, TechCrunch says, have generally resisted mandatory government oversight, favouring voluntary "embedded evaluators."

How the process would work

Reading the reported provisions together gives a sequence with several gates:

1. An evaluator registered with Commerce assesses a lab and its frontier models.
2. If it finds an imminent enumerated risk that meets the dollar or casualty thresholds, it has 72 hours to tell Commerce.
3. Commerce, not any other agency, may go to court.
4. A D.C. federal judge decides whether to order the model suspended.

Two features stand out. First, the thresholds are high and the standard is "imminent," which would narrow when anything could be blocked. Second, the sole-authority clause would stop other agencies from acting on their own. The Law Commentary site, which analysed the Politico report, points out that the draft comes from a company that would be regulated by it, which in its view makes the limits on government power worth scrutinising. It also asks how the independence of lab-paid evaluators would be checked, since the reported language only requires that payment not depend on results. TechCrunch raises the same question. These are commentators' questions, not findings.

What the bill would and would not cover

TechCrunch observes that the scenarios in the draft, such as bioweapons, loss of control and cyberattacks, are somewhat removed from the concerns that dominate the industry today, which it describes as agentic hacking, and it cites a run of recent incidents in which AI programs reportedly disobeyed operators. Whether the "evading its developer's controls" category would capture such incidents depends on the bill text, which we have not seen.

What it means

For AI labs, the practical effect would be a federal, court-gated intervention power in place of agency discretion. Labs that already accept outside evaluators, as Anthropic now does with METR, would have a head start; smaller developers might face new testing and evaluation costs. Punchbowl News reported that some Hill staff worry Anthropic's approach would create paperwork that blocks startups; that is one outlet's account of staff views and Anthropic has not endorsed that reading.

For lawmakers, the draft is one input among several. Punchbowl reports that Senate Majority Leader John Thune, Commerce Committee Chair Ted Cruz and Sen. Amy Klobuchar were close to proposing a bill on advanced AI risk, with a legal duty on AI firms to manage frontier risks in advance and to alert Commerce to major capability jumps, but that the deal is on hold amid disagreements with Anthropic and with Commerce Committee ranking member Maria Cantwell, which pushed the committee's markup until after recess. According to Punchbowl, Anthropic wanted to focus on its own obligations, called for public disclosure of risks, and objected to giving Commerce an ongoing injunction power. An Anthropic spokesperson said the company "continue[s] to work with the committee" on catastrophic risk. Cantwell, per the same report, wants negotiations to start from updates to bills the committee approved in 2024 and prefers that Commerce, national labs and national security experts lead testing.

A separate measure exists: Law Commentary reports that Sen. Mark Warner and colleagues introduced S. 5576 on 29 September 2026, creating an AI Safety Board with pre-release testing, and that Cruz objected when Warner sought unanimous consent. We have not verified that independently.

What we don't know yet

• The text. Nobody outside the drafting circle has published the bill, so definitions, penalties and procedure are unknown.
• Preemption. Some secondary outlets say the draft would override certain state AI laws on risk disclosure, third-party evaluation and incident reporting; the Reuters-based summary we read does not address it, and one outlet notes this would sit uneasily with Anthropic's June position that states should keep authority without a strong federal law. Treat it as unconfirmed.
• Judicial approval. Reports differ on whether a judge must approve a block; the Politico-based accounts we read say Commerce would ask a judge to halt a model.
• A sponsor and a schedule. No lawmaker has been named, and Congress is reported to be out of session until after the midterm elections.
• How the draft squares with the Punchbowl report that Anthropic objected to an ongoing Commerce injunction power. A judge-gated, case-by-case request is not the same thing as standing authority, but we cannot tell from the reporting.

What to watch next

Watch for a sponsor stepping forward, or for the text being published; for whether the Thune-Cruz-Klobuchar effort restarts after the recess and what it says about court action and preemption; for METR's first published assessment under the new access arrangement; and for how OpenAI, Google DeepMind and Nvidia respond to a lab-authored, court-gated proposal.

Sources

• Reuters via The Star: Anthropic quietly pitches its own AI safety bill to lawmakers, Politico reports (17 October 2026)
• TechCrunch: Anthropic pitches AI safety bill that may block a model if 50+ people could be killed (20 October 2026)
• Anthropic: An update on independent evaluators (15 October 2026)
• Punchbowl News: Thune, Anthropic and the stalled AI bill
• Law Commentary: Anthropic AI safety bill and government powers