1. What does Anthropic's AI safety bill propose?
As of October 26, 2026, Anthropic's draft federal AI safety bill is drawing criticism from safety advocates. The company, which develops the Claude family of models, has shown US lawmakers a proposal under which the government could halt a frontier AI model only when the danger is imminent and catastrophic. The text has not been formally introduced in Congress.
The proposal amounts to the developer's own answer to a question Washington has not settled: how much authority federal officials should hold over the most capable AI systems, and at what point they may use it. Because the author is one of the companies such a law would cover, the draft has attracted scrutiny well beyond its technical provisions.
2. A pitch, not a law
The draft was prepared in October 2026 and circulated among lawmakers as a confidential proposal. No member of Congress has put forward the text as a bill, which means it carries no legal force and may change substantially before anyone sponsors it. At this stage it works as a statement of the rules Anthropic would prefer to see adopted.
That distinction matters for reading the reaction. Much of the criticism targets what the document signals about industry priorities rather than any binding obligation, since nothing in it can yet be enforced. Whether it evolves into a vehicle for legislation depends on whether a lawmaker chooses to adopt it, amend it or set it aside in favor of a competing approach.
Drafts of this kind often serve as an opening position in a negotiation. By placing specific numbers and a specific agency on the table first, a company can shape the vocabulary of the debate, even if the final language looks very different. Critics of the proposal see that agenda-setting effect as its real purpose.
3. The thresholds that trigger government action
The core mechanism is a pair of numerical triggers. Under the draft, the government could block the release or use of an advanced model only when it identifies a foreseeable, imminent risk of more than $1 billion in damages. A second trigger covers human harm: intervention would be allowed if a model threatened to kill or seriously injure more than 50 people.
The text also confines the covered dangers to three scenarios. The first is a model helping someone create or release a biological weapon. The second is a model escaping the control of the company that built it. The third is cyberattacks. A risk falling outside those categories would not qualify, however large it might otherwise be.
Two qualifiers narrow the reach further. The risk must be foreseeable, meaning it can reasonably be anticipated, and it must be imminent, meaning it is close at hand. Together with the dollar and casualty figures, these conditions set a high bar that a model would have to clear before any intervention could begin.
The practical meaning of these figures is easiest to see by considering what falls beneath them. A model whose misuse could cause hundreds of millions of dollars in losses, or injure a few dozen people, would not meet either test, and the government would have no power under the draft to pause its deployment. Whether that outcome is acceptable is the question at the center of the dispute.
The three named categories also reflect a choice about what counts as a frontier risk. Biological weapons, loss of developer control and cyberattacks are the scenarios most often cited in discussions of catastrophic AI misuse. Other harms that regulators and researchers raise, such as large-scale fraud or manipulation, are not covered by the trigger as described.
4. Who holds the off switch?
Authority under the draft is concentrated in a single body. The Department of Commerce would be the only agency able to go to court to pause the deployment of a specific model. Other agencies would not have that power, so a sector regulator that saw a danger in its own field could not act independently.
The proposal also builds in outside testing. It calls for repeated independent evaluations of frontier models, and an evaluator that detects an imminent risk would have 72 hours to report it to Commerce. That deadline is meant to ensure a warning reaches the one agency empowered to act without lingering inside a company or a testing firm.
Beyond the evaluation regime, the draft contains whistleblower protections and a structure for reporting incidents. These provisions are intended to give employees and companies a defined channel for flagging problems, which is the information the government would need before it could meet any of the triggers described above.
The reliance on independent evaluators raises its own practical questions, including how evaluators are selected, how they are funded and how their findings are reviewed. The draft's 72-hour reporting window sets the pace for escalation, but the speed of any court proceeding that follows would determine how quickly a model could actually be paused.
5. How does the draft borrow from California's SB 53?
The proposal draws on California's SB 53, a state law signed in 2025 that deals with the risks posed by advanced AI models. It reflects that law's thresholds but sets a higher bar before federal action is possible. In effect, it takes a framework built for state-level oversight and adapts it to a national setting with a more demanding standard for stepping in.
Borrowing from an existing statute gives the draft a recognizable starting point for lawmakers who have already studied the California approach. It also gives critics an obvious comparison, because the question becomes why a federal measure would be more permissive about intervention than the state rules it draws upon.
The comparison also highlights a difference in purpose. A disclosure law asks companies to assess risks and report them, while a power to halt a model is a far stronger tool that restricts what a company may release. Setting the threshold for the stronger tool at a high level is consistent with treating it as a last resort, which is how supporters of the draft are likely to frame it.
6. Why are critics pushing back?
The sharpest objection concerns authorship. One opinion piece called the bill a fairly transparent attempt to calm public fears that Anthropic's own people helped stoke, and argued that governments, not AI companies, should write the rules. In that reading, the draft responds to worries held by the company's current and former staff and its executives that the systems it builds could cause catastrophic harm.
AI safety advocates have focused on the numbers. They describe the $1 billion and 50-person thresholds as too lenient, arguing that real harm could occur well below those levels and that waiting for a risk to be both foreseeable and imminent excludes plausible dangers that have not yet materialized. They argue that a stricter standard would better protect the public.
The combination of conditions is the heart of the complaint. A model would have to threaten a very large loss or mass casualties, fall into one of three named categories, and present a danger that is already close. Critics say that leaves little room for early action, which is the point at which regulation is usually most useful.
Supporters of a narrow trigger can point to the costs of broad authority. A low threshold could expose developers to repeated interruptions, and a power to pause a model on thin evidence might be contested in court or used unpredictably. The draft's design reflects an effort to reserve the power for extreme cases, though critics argue it has been reserved for cases so extreme that it may never be used.
7. Industry-written rules and the wider debate
At bottom, the dispute is about whether an industry should draft its own guardrails. The opinion piece answered firmly that it should not. Companies that sell a technology, it argued, are poorly placed to decide how much constraint that technology should face, and elected governments should regulate AI firms rather than the reverse.
Industry-drafted legislation is not new in Washington, where companies and trade groups routinely supply language to lawmakers. What stands out here is the position of the author. The company proposing lighter triggers is also among the loudest voices warning about catastrophic AI risk, an overlap that makes the draft harder to categorize as simple self-interest or simple caution.
The same tension is familiar to readers who follow digital assets, where firms have long lobbied for rules written to fit their business models while regulators and critics push for broader consumer and market protections. In both fields, the debate turns on who shapes the framework before it hardens into law.
For companies, investors and developers building on top of frontier models, the outcome has practical weight. The rules that emerge will influence how models are tested before release, which agency answers for failures and how much uncertainty businesses carry when they rely on third-party AI systems. That is why a confidential draft has drawn attention well beyond Capitol Hill.
8. What happens next for the draft
The draft is one option among several and would compete with other approaches to federal AI oversight. The contrast between them is likely to influence how strict any eventual standard becomes, since lawmakers will weigh a narrower, industry-backed trigger against calls for earlier and broader oversight.
The single-agency design carries its own trade-off. Giving Commerce exclusive authority could speed decisions by removing turf disputes, but it would also create a single point of failure if the department proved slow, under-resourced or reluctant to act.
Observers will watch whether Anthropic lowers its thresholds, widens the list of covered risks or only tidies the wording. The larger test is whether any member of Congress formally introduces the text. Until someone does, it remains a description of Anthropic's preferred rules, and that is precisely what its critics object to.
Any movement will probably come in stages. A sponsor would need to introduce text, committees would need to hold hearings, and competing proposals would need to be reconciled before a vote. At each step, the thresholds and the question of which agency holds the power to act are likely to remain the most contested elements.