Markdown edition

# Anthropic’s reported AI safety bill sets thresholds for blocking frontier models

> Reuters reports that Anthropic pitched lawmakers a private draft AI safety bill. Its reported harm thresholds differ from the developer-coverage thresholds in Anthropic’s public June framework.

By BIG CHANGE Editorial

Published: 2026-10-18T22:03:34.834Z
Updated: 2026-10-18T22:03:34.834Z
Canonical: https://bigchange.ai/blog/anthropic-reported-draft-ai-safety-bill-thresholds

![Conceptual illustration of an empty courtroom seen from its rear gallery, with counsel tables, a central aisle and an elevated judge’s bench.](https://bigchange.ai/api/media/file/anthropic-reported-safety-bill-courtroom-concept-v1.png)
AI-generated conceptual illustration by BIG CHANGE; it depicts no actual hearing, court order, jurisdiction-specific courthouse, or enacted bill.

Anthropic representatives have pitched US lawmakers on a draft bill for frontier AI safety and security, [Reuters reported on October 16, 2026](https://www.investing.com/news/stock-market-news/anthropic-quietly-pitches-its-own-ai-safety-bill-to-lawmakers-politico-reports-4954941), citing a Politico report based on three people familiar with the effort. The draft has circulated privately. Its text was unavailable for this article, so the reported provisions cannot be checked against the bill itself. The reported pitch is not an introduced bill or enacted law.

## The big change

- **What changed:** After publishing a policy framework in June, Anthropic is pitching lawmakers on a private draft with a reported test for stopping dangerous frontier models, according to Reuters. The test and the draft’s wording have yet to be verified from the document.
- **Why it matters:** The reported proposal assigns distinct roles to independent evaluators, Commerce and a federal judge. It would confine a model suspension to specified, imminent catastrophic harms, making the precise definition of those harms consequential for developers and the public.
- **What to watch:** A released draft or introduced bill would show whether the reported harm thresholds, Commerce’s role and limits on other agencies survive into actual legislative text. It would also show which parts of Anthropic’s wider June framework the bill includes.

The reported draft sets conditions for when the government could stop a model. According to Reuters’ account of Politico’s reporting, it would allow a block on release or use only if a model posed a foreseeable, imminent risk of more than $1 billion in damage or of killing or seriously injuring more than 50 people. Reuters said the risk would have to arise from creation or release of a biological weapon, a model escaping its developer’s controls, or a cyberattack. Those figures describe the reported intervention test.

## How the reported bill would work

The [Reuters report](https://www.investing.com/news/stock-market-news/anthropic-quietly-pitches-its-own-ai-safety-bill-to-lawmakers-politico-reports-4954941) says independent evaluators would register with the Commerce Department and periodically assess an AI lab and its frontier models. A lab could pay its evaluator as long as payment did not depend on the result. If an evaluator found what the report calls an “imminent enumerated risk,” the evaluator would have 72 hours after completing the assessment to notify Commerce.

Commerce could then ask a judge to halt a model after finding risk, according to Reuters’ account of the reported draft. Reuters says the proposed language would bar other US departments, agencies and officers from suspending a model’s development, deployment or internal use. A federal judge in the District Court for the District of Columbia could order a suspension. Without the draft text, the exact definitions, procedures and reach of those limits remain unverified.

An Anthropic spokesperson told Reuters that the bill represents proposals in the company’s [Advanced AI Framework](https://www-cdn.anthropic.com/files/4zrzovbb/website/0a58d567024a8b448ff15158ebc3625328dfcc1f.pdf), published in June 2026. The spokesperson said the proposal would let the government block or deter dangerous deployments of the most powerful models and require developers to test frontier models, disclose findings and obtain independent evaluations. Reuters said the Commerce Department did not immediately respond to its request for comment.

## What Anthropic’s public framework actually says

The [19-page June framework](https://www-cdn.anthropic.com/files/4zrzovbb/website/0a58d567024a8b448ff15158ebc3625328dfcc1f.pdf) proposes obligations for large frontier developers. Its suggested developer test has two parts: a model requiring more than 10^25 training FLOP, plus either more than $500 million in annual AI-derived revenue or more than $1 billion a year in AI research and development spending. The $1 billion spending figure in this public document concerns which developers would be covered. It is separate from the $1 billion damage figure attributed to the private draft.

The framework asks covered developers to assess biological weapons, offensive cyber operations, loss of control and automated research and development that could amplify those risks. It calls for published safety frameworks and system cards, risk reports at least every six months, and reports of critical safety incidents to a designated agency within 15 days of discovery. It proposes regular access for qualified independent evaluators to models and unredacted risk reports, with published evaluator reviews and conflict safeguards.

On enforcement, the framework proposes penalties for false statements and failures to evaluate, publish or report. It offers several possible ways to stop unsafe deployments, including an agency pursuing a court order, rather than prescribing one final design. Its discussion of government action includes safeguards against broad discretion and a route for judicial review. The document also recommends investments in biological and cyber resilience beyond developer obligations. Reuters’ account does not establish which of those recommendations, if any, appear in the circulated draft.

The draft’s precise language, any legislative sponsor, and treatment of issues such as state law remain unknown from the sources available here. Anthropic’s June framework argues for narrow limits on state-law preemption under a rigorous federal regime, but Reuters does not establish the draft’s position on that question.

## Sources & further reading

- [Reuters, “Anthropic quietly pitches its own AI safety bill to lawmakers, Politico reports” (October 16, 2026)](https://www.investing.com/news/stock-market-news/anthropic-quietly-pitches-its-own-ai-safety-bill-to-lawmakers-politico-reports-4954941). Reports Politico’s account, citing three people, of the privately circulated draft’s thresholds, evaluator role and court process; includes Anthropic’s response and notes that Commerce did not immediately respond. Reuters does not provide the draft text.
- [Anthropic, *Advanced AI Framework* (June 2026)](https://www-cdn.anthropic.com/files/4zrzovbb/website/0a58d567024a8b448ff15158ebc3625328dfcc1f.pdf). The public primary document for Anthropic’s proposed developer scope, evaluations, disclosures, enforcement options and resilience measures. It does not authenticate the language of the reported private draft.

## Sources

- [Reuters, Anthropic quietly pitches its own AI safety bill to lawmakers, Politico reports (2026-10-16)](https://www.investing.com/news/stock-market-news/anthropic-quietly-pitches-its-own-ai-safety-bill-to-lawmakers-politico-reports-4954941) — Reuters' October 16 account of Politico reporting based on three people. It reports the draft's harm thresholds, evaluator and Commerce roles, possible judicial suspension, Anthropic's spokesperson response, and Commerce's lack of immediate response. The private draft was not retrieved, so its text and scope remain unverified.
- [Anthropic, Advanced AI Framework (2026-06)](https://www-cdn.anthropic.com/files/4zrzovbb/website/0a58d567024a8b448ff15158ebc3625328dfcc1f.pdf) — Anthropic's public June 2026 proposal: developer-coverage criteria, four risk categories, transparency and independent evaluation, enforcement options, safeguards, and societal resilience. Its annual AI R&D spending figure is separate from the harm threshold Reuters attributes to the private draft. The PDF cannot authenticate that draft.