AI SAFETY: Anthropic Bill 2026: Its $1B Kill-Switch Rule

By Daniel Rao | October 17, 2026

Somebody has to write America's AI safety law. This week we learned who volunteered for the job: Anthropic, one of the companies the law would regulate.

According to Politico reporting published Friday, Anthropic representatives have been quietly pitching Capitol Hill lawmakers a draft bill to regulate the safety of frontier AI. Three people with knowledge of the effort confirmed the pitch to Politico, and both Anthropic and the Commerce Department declined to comment when Reuters asked about it.

The pitch landed in the same week Trump's new AI task force held its first meeting and the FBI warned of AI-powered election trickery before the midterms. Washington is having its AI week — and the company that builds the models wants to write the rules before the government writes them alone.

The $1 billion bar — and why almost nothing clears it

The draft's centerpiece is its trigger: the government could 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.

Read that twice. Under Anthropic's draft, a model could be biased, hallucinate dangerous financial advice, displace workers, or flood the internet with spam — and the kill switch stays in its case. Only three specific horrors qualify: the model helping create or release a biological weapon, the model escaping its developer's control, or a cyberattack.

The formula isn't invented from thin air. A bipartisan draft that circulated in Congress this summer used the same catastrophic-risk definition — $1 billion in damage or 50-plus casualties. Anthropic's version adopts the numbers but keeps the enforcement door as narrow as possible: nothing but an imminent, enumerated catastrophe opens it.

One judge in D.C., and a lockout for everyone else

Even when the bar is met, stopping a model is a two-key operation. Independent evaluators — private firms registered with the Commerce Department — would periodically test frontier labs' models. If they find an "imminent enumerated risk," they must alert Commerce within 72 hours. The labs themselves pay the evaluators, though the bill bars tying payment to the outcome of the evaluation.

Then only Commerce can go to court: it may ask a federal judge in the District of Columbia to order a model's suspension. And here is the clause that reveals the draft's deeper intent — the language explicitly prohibits any other US government department, agency, or officer from suspending a model's development, deployment, or internal use. Not the FTC, not the Pentagon, not the new AI task force. One department, one courthouse, one key.

For a lab that ships some of the world's most capable models, that clause is the real prize: a federal shield that decides, in advance, who is never allowed to pull the plug.

The rival bill with

The rival bill with a bar ten times lower

This is where the numbers get interesting. At this week's CEPA Tech & Security Conference, lawmakers from both parties laid out their own visions of AI safety — and the contrast is brutal.

The Kill Switch Act from Rep. Nathaniel Moran (R-Texas) and Rep. Ted Lieu (D-California) would require frontier developers to be able to throttle, suspend, or completely shut down a model that threatens $100 million in economic damage or 10 deaths — a bar ten times lower on money and five times lower on lives than Anthropic's draft.

The bipartisan FRONTIER Act from Rep. Jay Obernolte (R-California) and Rep. Lori Trahan (D-Massachusetts) takes a different route: no kill switch at all, but mandatory transparency reports, published risk-management frameworks, incident reporting, and audits by accredited independent verification organizations — plus one uniform federal standard that would override state rules such as California's. Emergency Commerce orders under the bill carry penalties of up to $1 million per violation, and $10 million for defying an emergency order.

"The brakes allow you to drive faster, not slower. AI brakes should not slow innovation. They should enable faster movement with a safe backstop." — Rep. Nathaniel Moran (R-TX), at the CEPA Tech & Security Conference

Congress is not short of AI bills — Senators Cruz, Thune, and Klobuchar are drafting catastrophic-risk legislation of their own. The open question is whose bar becomes law — Anthropic's near-impossible one, or the lawmakers' stricter ones.

Why the pitch is happening right now

Anthropic's quiet lobbying is racing a clock. On Thursday, President Trump's Super Intelligence Task Force held its inaugural meeting — a 120-day body led by Director of National Intelligence Jay Clayton as AI czar, tasked with commissioning a report on AI's dangers and setting up protocols for early warnings from frontier labs on breaches, hacks, and jailbreaks. Clayton told CNBC the meeting was "very good" and built on the White House Accord Trump signed with AI leaders last month.

An AI lab getting its own draft into congressional offices days before that task force starts writing reports is not a coincidence. It is the oldest move in Washington: get your version of the rules into the room first, before the room writes its own.

And the week piled on the pressure to act: on Friday, the FBI and CISA jointly warned election officials that hackers are using AI to sharpen phishing attacks ahead of next month's midterms, while the Wesleyan Media Project reported campaigns have spent $137 million on AI-generated political ads this cycle — only 27% disclosing the AI.

What it means for you

Nothing in Anthropic's draft would change how you use AI tomorrow. This is a bill about the most powerful models at a handful of frontier labs — OpenAI, Anthropic, Google DeepMind, Meta, xAI — not about the apps on your phone.

What it could change is how fast the most capable models reach you, and who gets to say no. A $1 billion / 50-person bar means almost no model would ever be stopped — which is exactly the point for a lab whose business is shipping. If Congress instead adopts the Kill Switch Act's $100-million threshold, or the FRONTIER Act's audit regime, frontier AI development would look materially different: slower releases, public risk reports, auditors inside the labs.

For voters, the through-line is simpler. The midterms are weeks away, AI-generated ads are already a nine-figure industry with minimal disclosure, and the federal government still has no comprehensive AI law. Whoever writes the first one sets the template for everything after — which is precisely why the company building the models is at the table.

The honest verdict

Anthropic drafting AI safety rules is not proof of bad faith. The company has spent years publicly calling for exactly this kind of regulation, and a government kill switch for genuinely catastrophic models is a reasonable idea — the industry's biggest names, from Anthropic to OpenAI, have all said some version of "please regulate us."

But the fine print reads less like a brake and more like a shield: a trigger so high it may never fire, a kill switch only one judge in Washington can reach, and a clause walling off every other federal enforcer. That is not regulation of the industry. That is the industry pre-regulating itself, on its own terms.

Congress now has at least three competing visions of AI safety on the table — and the week's task-force meetings, FBI warnings, and midterm pressure guarantee a fourth and fifth are coming. Watch which bar survives. Because the number lawmakers pick will decide whether "AI safety law" means anything at all.

The three bills, side by side

	Anthropic draft	Kill Switch Act (Moran/Lieu)	FRONTIER Act (Obernolte/Trahan)
Trigger to halt	Imminent risk of $1B+ damage or 50+ deaths/injuries	Threat of $100M+ damage or 10 deaths	No kill switch
How it stops	Commerce asks a D.C. federal judge	Developers must be able to throttle/suspend/shut down	Emergency Commerce orders; up to $10M penalties
Oversight	Independent evaluators, Commerce-registered, lab-funded	—	Accredited independent verification orgs; transparency reports
Status	Draft pitch, not introduced	Proposal	Introduced, in committee

Frequently asked questions

What exactly did Anthropic pitch to Congress?
A draft bill regulating frontier AI safety and security, pitched quietly to Capitol Hill lawmakers this week, according to Politico's reporting citing three people familiar with the effort. It has not been introduced as legislation, and Anthropic declined to comment.

Could the government shut down an AI chatbot under this bill?
Almost certainly not. The halt power applies only to imminent risks of $1 billion-plus in damage or 50-plus deaths or serious injuries, tied to bioweapons, loss of model control, or cyberattacks. Ordinary consumer models would not qualify.

How does this differ from the FRONTIER Act?
The FRONTIER Act is a bipartisan transparency-and-audit bill: public risk frameworks, transparency reports, accredited third-party auditors, and a uniform federal standard overriding state rules — with no kill switch. Anthropic's draft is narrower: a halt mechanism only, set at an extreme bar, plus a clause barring other federal agencies from suspending models.

When could any AI safety bill become law?
Not soon. None of these drafts has cleared committee, the House was on recess this week and returns next, and the midterm elections will reshape the calendar. The pressure is building — the FBI's midterm warning and the task force's 120-day clock are real — but the law is still unwritten.

Sources

Published October 17, 2026 · By Daniel Rao

Search This Blog

TechPulse USA