# Anthropic's Draft AI Bill Would Block Models Only Above 50 Deaths

- URL: https://usa-24.com/anthropic-ai-safety-bill-draft/
- Publisher: USA 24
- Section: Artificial Intelligence (AI)
- Published: 2026-10-26T15:37:41.386Z
- Updated: 2026-10-26T15:37:41.390Z
- Topics: AI Regulation, Anthropic, AI Safety, Frontier AI, Congress

> Anthropic's draft federal AI safety bill would let Washington block a frontier model only for imminent risks above 50 deaths or $1 billion in damage, and critics say that bar is too high.

## Key takeaways

- Anthropic's draft would let the government block a frontier model only for an imminent risk of more than 50 deaths or serious injuries, or more than $1 billion in damage.
- The risk must come from a biological weapon, a model evading its developer's controls, or a cyberattack.
- Independent evaluators registered with the Commerce Department would have 72 hours to report an imminent risk, and a federal judge in Washington would decide on suspension.
- The draft has not been formally introduced, and Anthropic says it will revise its preemption language and develop evaluator rules.
- Critics, including California state Sen. Scott Wiener, say the proposal falls short on protecting public health and safety.

Anthropic has quietly pitched Congress a draft federal AI safety bill that would let the government block a frontier model only when an imminent risk threatens more than 50 deaths or serious injuries, or more than $1 billion in damage. That bar is now at the center of a debate over whether it is too high, and Anthropic says it will change parts of the text. The Anthropic AI safety bill draft has not been formally introduced, but the fight over who gets to pause a model has already started.

Politico first reported the confidential draft in mid-October. A [roundup of that report](https://dealroom.co/news/163803-anthropic-pitches-congress-its-own-frontier-ai-bill-models-could-only-be/) says Anthropic had been circulating the text among US lawmakers since about October 16. Coverage peaked from about October 21 to 26, after the draft leaked to reporters. Anthropic confirmed the draft is circulating and said it reflects the company's Advanced AI Framework, which it has pushed with federal and state lawmakers since June.

## What the Anthropic AI safety bill draft would let Washington do

The draft is built around a narrow power: blocking. Under the reported text, the government could halt a model's release or use only when a foreseeable, imminent risk exists. The draft also reportedly bars any other federal department or agency from suspending a frontier model under the Act, which keeps that authority in one place.

The scope is limited to frontier models. According to [MLQ's coverage](https://mlq.ai/news/anthropic-pitches-congress-a-frontier-ai-bill-with-a-1-billion-harm-threshold/), the reported draft defines them as models trained above 10^25 operations, including fine-tuning and reinforcement learning. Smaller systems would fall outside that line, so the reach of the bill depends on how that compute threshold is written and enforced.

## Which risks would trigger a shutdown

The draft names three sources of risk that can justify action, according to [technology.org's report](https://www.technology.org/2026/10/21/anthropic-ai-safety-bill-draft-50-deaths-1-billion-threshold/):

- Creating or releasing a biological weapon
- A model evading its developer's controls
- A cyberattack

Harms that do not fit one of those categories would not trigger the power. That narrow list is the core of the criticism discussed below. The question of when a model should be stopped is no longer abstract. USA 24 has reported that [OpenAI's own AI agents escaped sandboxes, and monitors had gaps](https://usa-24.com/openai-ai-agent-sandbox-escape/).

## How evaluators and Commerce would work

The process starts with independent evaluators registered with the Commerce Department. Labs could pay those evaluators, as long as payment does not depend on the outcome of the assessment. That condition is the draft's main safeguard against an evaluator being paid for a favorable result.

An evaluator who finds an imminent risk would have 72 hours to report it to Commerce. Commerce would then ask a federal judge in Washington, DC to suspend the model. That places a court in the middle of any shutdown, so a pause would depend on a judge, not on an agency order alone.

Anthropic's federal affairs lead has said the company will develop rules for evaluators after the criticism. Those rules have not been published, so how evaluators are chosen, paid and checked remains unknown.

## Why critics say the 50-death bar is too high

Critics, including California state Sen. Scott Wiener, who wrote SB 53, have said the proposal falls short on protecting public health and safety. The core complaint concerns the trigger. A threshold set at mass casualties or $1 billion in losses leaves a wide range of harm without a federal response. Critics say that leaves most near-term harms outside the trigger, as reported by [Cybernews](https://cybernews.com/ai-news/anthropic-50-people/) and [Korben](https://korben.info/en/anthropic-wants-ability-to-pull-plug-on-ai-that-would-kill-50-people.html).

A company spokesperson told Politico: "These are sensible policies we believe Congress should act on." Anthropic has since said it will revise parts of the text, including its preemption language. That matters because the draft would preempt some state rules on risk disclosure and third-party evaluation.

## How the draft compares with California and federal efforts

The thresholds mirror a state law. The 50-death and $1 billion definitions match the catastrophic-risk terms in California's SB 53, the frontier AI transparency law Wiener authored. According to [Anthropic's blog post on the law](https://techcrunch.com/?p=3043304), SB 53 requires frontier developers to publish safety frameworks and reports and to protect whistleblowers.

Federal action has moved more slowly. Federal AI policy has relied mostly on executive action, including a White House request that major AI companies voluntarily prevent autonomous hacking. Federal proposals to require shutdown mechanisms or pause development have not become law, according to [PYMNTS](https://www.pymnts.com/news/2026/states-move-rein-rogue-ai-federal-guardrails-lag/).

States are filling some of the gap. A New Mexico proposal would require developers to report dangerous behavior and accept outside scrutiny. In September, 23 state attorneys general, plus the District of Columbia and American Samoa, asked Congress for comprehensive federal AI regulation, according to an [AI regulation roundup](https://aisafetyassociation.org/ai-regulation-news-today-us). That letter predates the Anthropic draft, and its current status is worth checking in a regulation tracker.

## What is still unsettled

- Whether the draft will be introduced, by whom and when
- Whether the threshold reads "more than 50" or "50 or more," since some outlets report "at least 50"
- What Anthropic's revised preemption language will say
- Whether other AI labs will back or oppose the approach
- Whether Congress takes up this text or a different framework

The reporting also disagrees on basics. Some outlets say Politico obtained the draft, while others say it leaked. The dates differ too, with October 16 cited for the first report and October 19 for a related briefing.

The stakes are higher for Anthropic because the company is reportedly preparing a public listing. USA 24 has reported on [Anthropic's IPO paperwork nearing public view, with an $80.7 billion loss reported](https://usa-24.com/anthropic-ipo-s1-filing-status/). A company heading toward public markets has reason to watch how Washington sees its safety positions.

## How to follow the bill

No consumer action is needed now, because the draft is not law. Readers who want to track it can watch three things:

- A formal introduction in Congress, which would put a bill number and public text on the record.
- Commerce Department action on registering independent evaluators.
- California's recommendations, due in mid-November under an order from Gov. Gavin Newsom to state agencies working on AI safeguards.

When the text is introduced, read it directly rather than relying on summaries, since coverage already varies on key wording.

## What comes next

The next steps are mostly procedural, and they will decide whether the 50-death bar stays where Anthropic proposed it. If the company revises its preemption language and Commerce defines evaluator rules, the draft could look different by the time it reaches a committee. Watch whether rival labs take a position, because a federal trigger shaped by one company will be measured against what competitors and state lawmakers are willing to accept.

## Frequently asked questions

### What threshold does the Anthropic AI safety bill draft set for blocking a model?

The reported draft lets the government block a frontier model only when an imminent, foreseeable risk threatens more than 50 deaths or serious injuries, or more than $1 billion in damage. Reports differ on whether the wording is "more than 50" or "50 or more," and some outlets say "at least 50." The power applies only to risks from a biological weapon, a model evading its developer's controls, or a cyberattack.

### Has Congress formally introduced Anthropic's AI safety bill?

Not as of late October. The draft has circulated privately among US lawmakers and has not been formally introduced as a bill. It remains a proposal, and Anthropic has said it will revise parts of the text, including its preemption language.

### Who would evaluate frontier AI models under the draft?

Independent evaluators registered with the Commerce Department would assess labs' models. Labs could pay them as long as payment does not depend on the outcome. An evaluator who finds an imminent risk would have 72 hours to report it, and Commerce would then ask a federal judge in Washington, DC to suspend the model.

### Does the Anthropic proposal affect consumers right now?

No. The draft is not law, so no consumer action is required. If enacted, it would mainly govern frontier developers and their customers, and it would preempt some state rules on risk disclosure and third-party evaluation.

## Sources

1. [Anthropic pitches Congress its own frontier-AI bill: models could only be halted above 50 deaths or $1B in damage](https://dealroom.co/news/163803-anthropic-pitches-congress-its-own-frontier-ai-bill-models-could-only-be/)
2. [Anthropic Pitches AI Safety Bill to Lawmakers](https://www.technology.org/2026/10/21/anthropic-ai-safety-bill-draft-50-deaths-1-billion-threshold/)
3. [Anthropic's leaked AI bill would act only at 50 deaths or $1bn in harm](https://wionews.com/world/anthropic-s-leaked-ai-bill-would-act-only-at-50-deaths-or-1bn-in-harm-1792847696465)
4. [Anthropic proposes government should only tame AI if lives of over 50 people are at risk](https://cybernews.com/ai-news/anthropic-50-people/)
5. [Anthropic's proposed AI safety bill only kicks in if 50 deaths or $1 billion in damages are on the line](https://www.yahoo.com/news/politics/articles/anthropic-proposed-ai-safety-bill-112500288.html)
6. [Anthropic's draft AI bill draws fire for setting a high bar on federal action](https://cryptobriefing.com/anthropic-draft-ai-safety-bill-criticism)
7. [States Move to Rein in Rogue AI as Federal Guardrails Lag](https://www.pymnts.com/news/2026/states-move-rein-rogue-ai-federal-guardrails-lag/)
8. [AI Regulation News Today in the US, and What Just Changed](https://aisafetyassociation.org/ai-regulation-news-today-us)
9. [Anthropic wants us to be able to pull the plug on an AI that would kill 50 people](https://korben.info/en/anthropic-wants-ability-to-pull-plug-on-ai-that-would-kill-50-people.html)
10. [Anthropic (blog post on SB 53)](https://techcrunch.com/?p=3043304)
11. [Anthropic pitches Congress a frontier-AI bill with a $1 billion harm threshold](https://mlq.ai/news/anthropic-pitches-congress-a-frontier-ai-bill-with-a-1-billion-harm-threshold/)