Anthropic has quietly put a confidential draft AI safety bill in front of Capitol Hill lawmakers, and the move is already setting off a fierce policy clash over how far Washington should go to police frontier models. POLITICO first reported the draft, saying it obtained the document and spoke with three people familiar with the effort. The proposal would give the government limited power to block a model’s release or use only in narrow cases tied to a foreseeable, imminent risk of more than $1 billion in damages or death or serious injury to more than 50 people. Anthropic says the draft reflects its own safety framework. Critics, at least in the emerging debate, now have a ready-made target.
What makes the document especially combustible is not just that it exists, but how tightly it tries to define when federal intervention would be allowed. According to POLITICO, the draft would only permit a suspension if the danger came from one of three sources: a biological weapon, a model escaping developer control, or a cyberattack. That is a far cry from a sweeping shutdown power. It is also a reminder that the AI policy fight in Washington is shifting from abstract principles to hard thresholds, legal triggers and who gets to pull the lever when a model crosses them.
For all the headline heat, this is still a confidential, un-introduced draft with no effective date. No lawmaker has introduced it, according to AInvest. That matters. The bill is not a law, not a formal committee proposal and not yet a public legislative vehicle with a sponsor list, markup schedule or vote count. It is a private pitch, which means the immediate debate is about influence and intent, not implementation. Even so, the fact that Anthropic went directly to lawmakers with a detailed framework suggests the company wants to shape the rules before Congress settles on a broader AI regime.
Anthropic confirmed on Friday, Oct. 16, 2026, that it is circulating the draft, saying it tracks its Advanced AI Framework that the company has promoted since June. One Anthropic spokesperson told POLITICO: “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. These are sensible policies we believe Congress should act on.” In other words, the company is not hiding from the substance of the plan. It is openly trying to turn its internal safety doctrine into statutory language.
The draft goes beyond emergency suspension powers. POLITICO reported that it would preempt some state AI laws, including rules on risk disclosure, third-party evaluation and incident reporting. At the same time, it would require safety frameworks, transparency reports, whistleblower protections and incident reporting. That combination is politically tricky: the bill asks Washington to standardize the highest-stakes rules while limiting what states can do on their own. For supporters, that could reduce a patchwork of conflicting requirements. For opponents, it looks like a federal override that narrows state authority while preserving a central gatekeeping role in the nation’s biggest AI fights.
The draft also specifies who can act. According to POLITICO, only the Commerce Department could seek a suspension, and only a federal judge for the D.C. District Court could order it. The draft would bar any other U.S. department, agency or officer from suspending a model. That restriction is crucial. It means the proposal is not a blank check for regulators and not a generalized emergency tool. It channels power through one cabinet department and one court, a structure that appears designed to reassure industry while still giving the government a way to step in if the worst-case scenario materializes.
The controversy is less about whether frontier AI should be regulated than about who gets to define danger and when. Anthropic’s proposal uses very high damage and casualty thresholds, which suggests the company wants a narrow legal backstop for extreme failures rather than a broad licensing system. That design may appeal to lawmakers looking for something concrete, but it also invites skepticism from those who see any shutdown authority as an opening for overreach, political abuse or regulatory creep. The phrase “AI kill switch” may be catchy, but the actual draft is more specific — and that precision may be exactly why it is causing a stir.
Anthropic’s own framing shows the company is trying to present the bill as a safety compromise rather than a power grab. In a separate quote carried by AFP via Le Devoir, a spokesperson said: “Our proposal gives the federal government the ability to block or deter dangerous deployments of the most powerful models.” That language is important. It does not promise routine intervention. It emphasizes deterrence and blocking in extreme cases. Still, once a company invites government into the release decision for the most advanced systems, the political fight becomes inevitable. The issue is no longer just technical risk. It is public authority.
The policy push lands at a sensitive corporate moment. Anthropic is a private, pre-IPO company, and investors are reportedly weighing a valuation near $1.5 trillion for a planned November IPO, down from a $2 trillion target, according to Reuters as cited by AInvest. Because Anthropic is not publicly traded, there is no market price to react in real time. But the scale of the reported valuation underscores how closely the company’s safety posture and regulatory ambitions are tied to its market story. A higher-profile governance role can reinforce a brand of discipline and seriousness, even as it raises questions about whether the company is trying to write the rules around its own future business.
That timing also matters because there is no immediate legislative path for the draft. AInvest said Congress is out of session until after the November 2026 midterm elections, which means no near-term action is imminent. So the bill’s real short-term effect is reputational and strategic. It puts Anthropic in the center of the policy conversation while lawmakers are away, and it gives the company a chance to test how much support there is for a federal safety framework that could preempt some state rules.
For now, the proposal is an opening salvo in a much larger fight over frontier AI governance. The draft’s thresholds, its narrow list of triggering risks and its limited federal chain of command all point to a company trying to make regulation predictable rather than expansive. But the very act of privately circulating a shutdown bill guarantees pushback. If lawmakers like the concept, they may rewrite it. If they dislike it, they may use it as evidence that industry wants to shape the guardrails on its own terms.
The next concrete milestone is not a hearing or a markup. It is Anthropic’s planned November IPO, where the reported $1.5 trillion valuation estimate will be weighed by investors. Until then, the bill will live in the gray zone between private influence and public law — exactly the kind of space where Washington’s AI fight is getting most intense.
