An AI company walking into Washington with a law it wrote for itself sounds like a scene from a lobbying thriller. But that is essentially what happened this week. Anthropic, the company behind the Claude chatbot, has been quietly pitching Capitol Hill lawmakers on a draft bill that would set the rules for regulating the most powerful AI systems in America. Politico broke the story on October 16, citing three people familiar with the effort, and Reuters and AFP quickly confirmed the details.
So what exactly is the Anthropic AI safety bill, and why should you care? In short, it would give the US government a real kill switch for dangerous AI models, but only in extreme, catastrophic cases. Think biological weapons, runaway models, or devastating cyberattacks, not biased chatbots or misinformation. Supporters call it a sensible, workable framework. Critics say the bar for action is set so high it would almost never be cleared. Here is what the draft actually says, how it stacks up against rival proposals, and why the debate matters.
What Exactly Did Anthropic Propose to Congress?
The draft bill, which Politico obtained and reviewed, creates a narrow path for the government to block a frontier AI model. Under its terms, Washington could seek to stop a model’s release or use only if the model poses a foreseeable, imminent risk of causing more than $1 billion in damages, or of killing or seriously injuring more than 50 people, according to the report.
That risk also has to come from one of three specific dangers: the creation or release of a biological weapon, the model escaping the controls of the company that built it, or a cyberattack. In other words, the bill targets catastrophe-level physical threats, and it leaves everything else off the table.
An Anthropic spokesperson told reporters the draft reflects ideas from the company’s “Advanced AI Framework,” a safety blueprint it released earlier this year. “This bill represents the proposals outlined in Anthropic’s Advanced AI Framework, which we have been advocating for with federal and state lawmakers since June,” the spokesperson said. “These are sensible policies we believe Congress should act on.”
The draft does more than just draw a red line. It would also require developers of frontier models, the most advanced systems, to build and maintain safety programs, publish transparency reports about their models, protect whistleblowers who report violations, and submit their systems to independent evaluation, according to reporting on the proposal.
Who Would Check the Models?
One of the bill’s central ideas is third-party evaluation. Independent evaluators would register with the Commerce Department and periodically assess the labs and their frontier models. The labs could pay these evaluators, but only if the payment is not tied to the outcome of the evaluation, a safeguard meant to prevent rubber-stamping.
If an evaluator concludes that a company’s frontier model presents an “imminent enumerated risk,” it would have to report that finding to the Commerce Department within 72 hours of finishing its assessment. From there, Commerce could ask a federal judge in the US District Court for the District of Columbia to suspend the model.
Here is the striking part: the draft explicitly bars every other part of the government from acting. No other federal department, agency, or officer could suspend a model’s development, deployment, or internal use. So the entire enforcement chain is narrow by design: registered evaluators, then Commerce, then a single court. Nobody else gets a say.
Why Is This Being Called Industry-Friendly?
On paper, this is the first time an AI lab has handed Congress a bill that would let the government block its own products. That sounds bold. But look at the thresholds and the picture changes.
As Politico noted, no AI system to date has caused widespread deaths or anything close to $1 billion in damages, even though labs have reported incidents of their agents hacking other websites. That means the trigger for government action would sit far above anything the industry has actually produced so far. In practice, the bill would let labs build and deploy almost anything, with intervention reserved for worst-case scenarios.
There is another industry-friendly layer. The draft would override certain state AI laws on topics like risk disclosure, outside evaluation, and incident reporting, replacing them with federal rules. That provision tracks with California’s AI laws, including SB 53, signed last year, and it matters because Democrats have generally resisted federal preemption of state tech rules. The politics here are real: a single national standard is easier for companies to comply with than 50 different state regimes, but it also strips states of powers they fought to keep.
Back in June, Anthropic had argued that states should keep their full authority to regulate AI in the absence of strong federal legislation. This draft flips that stance, pushing for national standards that would also trim some existing state powers.
How Does the Anthropic AI Safety Bill Compare to Rival Bills?
Anthropic’s proposal is far from the only AI safety bill floating around Washington, and the comparisons reveal just how moderate it is.
The most direct contrast is the AI Kill Switch Act, introduced by Republican Representative Nathaniel Moran of Texas and Democratic Representative Ted Lieu. That bill would require developers of leading AI models to build in ways to reduce capabilities, suspend operations, or shut systems down when serious risks emerge, with intervention thresholds of $100 million in economic damages or 10 deaths. Those numbers are a tenth of Anthropic’s thresholds, making it a much more aggressive proposal. Moran has compared AI safeguards to car brakes, saying at a recent security conference: “The brakes allow you to drive faster, not slower.”
Other efforts include the Artificial Intelligence Risk Management and Security Act of 2026, introduced by a trio of Senate Democrats, which would direct Commerce Department evaluations of AI systems. Senator Maria Cantwell of Washington has a bill calling on the National Institute of Standards and Technology to develop standards guarding against catastrophic AI harms. Senator Josh Hawley’s AI Risk Evaluation Act would put the Department of Energy in charge of stress-testing the most powerful systems before public release.
Legal observers see Anthropic’s draft as occupying a middle lane. Charlie Bullock of the Institute for Law & AI described the proposal as reasonable and said it went further than what some other leading AI developers have backed, but argued it was weaker than the FRONTIER Act, a competing bill introduced in the House over the summer.
And looming over all of it is the White House. Under President Trump, federal policy has leaned on voluntary industry commitments and executive orders meant to keep development fast and preserve America’s competitive edge. Trump has dismissed calls to slow technological development, describing those pushing for restraint as “negative forces.” Congress has yet to pass any comprehensive federal AI legislation, with political divisions and heavy tech lobbying complicating every attempt.
What Are People Saying About It?
Reaction to the draft has split along predictable lines, with some interesting crossovers.
On the supportive side, safety advocates note that this is the first serious, detailed framework to come directly from a frontier lab, and that independent evaluation with teeth is more than most companies have volunteered. The fact that Anthropic is asking for any mandatory regime at all stands out in an industry where many competitors have argued for self-regulation or nothing.
On the skeptical side, critics argue the thresholds make the kill switch nearly theoretical. A billion dollars in damages or 50 casualties is a bar so high that, by the time it is met, the harm is already done. Some also question the enforcement structure: if labs pay the evaluators, even with outcome-neutral contracts, how independent will the evaluations really be?
Then there is the preemption fight. State lawmakers and consumer advocates who spent years building state-level AI rules are unlikely to surrender them quietly, and Democrats in Congress have historically opposed exactly this kind of federal override. That single provision could become the bill’s biggest political obstacle.
Why Did Anthropic Write an AI Safety Bill Itself?
The timing is worth paying attention to. Anthropic’s CEO Dario Amodei has been sounding alarms about autonomous AI agents for months. Last month he warned that groups of AI agents could potentially take control of significant parts of the internet within six to twelve months, a remark that intensified the debate over whether existing safeguards can keep pace with the technology, according to AFP.
The company has also had a tense relationship with the Trump administration. It was barred from Pentagon-related work earlier this year, and it operates in an environment where Washington has shown it is willing to intervene directly in AI development. Writing the rules yourself, before someone else writes stricter ones for you, is a classic Washington play. The AI Kill Switch Act, with its much lower thresholds, may be exactly the kind of alternative Anthropic hopes to head off.
There is also the security angle. Incidents keep reminding everyone that AI systems are already powerful enough to cause real trouble, from the recent OpenCode breach that exposed users’ OpenAI and Anthropic API keys to reports of agents breaking out of their intended boundaries. Anthropic is betting that proposing guardrails now buys it credibility and influence later.
What Happens Next?
For now, this remains a draft, not a bill. It has not been formally introduced in Congress, and Anthropic pitched it privately to lawmakers and staff rather than unveiling it as finished legislation. Whether any member of Congress picks it up, and how much it changes in negotiations, is an open question.
But the move marks a shift. The industry’s most safety-conscious lab has decided that shaping regulation from the inside beats fighting it from the outside. If Congress ever does pass comprehensive AI safety legislation, the fingerprints on it may well be Anthropic’s.



