Anthropic's proposed federal AI safety bill would allow authorities to seek suspension of an advanced AI model if it posed a foreseeable, imminent risk of killing or seriously injuring more than 50 people or causing over $1 billion in damages. The government could seek intervention before the harm occurred.
According to an October 16 Reuters report citing Politico, Anthropic presented the proposal to congressional lawmakers. The draft has not become federal law, and exceeding the proposed threshold would provide grounds to seek a court order rather than automatically prohibit a model.
The proposal raises questions about how quickly officials could respond to emerging threats and which interventions would remain available when anticipated harm falls below the numerical thresholds.
How Commerce could seek to suspend an AI model
Under the reported proposal, independent evaluators registered with the Commerce Department would regularly review advanced AI developers and their models for catastrophic risks. Assessments would cover potential assistance with biological weapons or cyberattacks. Evaluators would also examine whether models could evade their developers’ controls.
If an assessment identified an imminent risk covered by the proposal, evaluators would have 72 hours after completing the review to notify Commerce. The department could then ask a federal judge in Washington, D.C., to suspend the model. Other federal departments and officials would face restrictions on independently halting model development or deployment, including use within the developer’s own operations.
Judicial review would give developers an opportunity to challenge the evidence supporting a suspension. However, requiring requests to pass through Commerce could delay action if another agency detected an urgent threat but could not intervene independently.
The complete draft has not been independently verified, so possible exceptions remain unclear. Its relationship with existing federal powers would also require examination of the legislative text.
What happens when an AI threat falls below the threshold?
A model capable of facilitating cyberattacks could threaten sensitive systems even when projected damages fall below $1 billion. Similarly, evidence that a model can bypass developer safeguards might justify additional testing or deployment restrictions before officials can establish an imminent catastrophic threat. These hypothetical scenarios illustrate risks that may not meet the proposed suspension criteria.
Congress would need to determine which interventions remain available below the suspension thresholds. Requirements for additional testing or corrective measures could address emerging risks before they justify stopping an entire model.
The proposal's reporting deadline raises a related timing question. Since the 72-hour period begins after an assessment is completed, lawmakers would also need to establish how evaluators should report urgent findings discovered during an unfinished review.
Developer-funded evaluations raise independence concerns
According to TechCrunch, developers could pay independent evaluators if compensation did not depend on findings. Although this condition would prohibit direct financial rewards for favorable conclusions, evaluators could still have incentives to retain developers as future clients.
Anthropic's published June Advanced AI Framework proposes safeguards such as pooled funding and government financing, along with conflict-of-interest disclosures. It also discusses randomly assigning qualified evaluators in high-stakes cases to reduce developers' ability to select reviewers based on their expected conclusions. Whether these safeguards appear in the draft legislation remains unverified.
Evaluator independence would also depend on access to the evidence needed for assessment and permission to report findings directly to regulators. As The Neuron's earlier coverage of independent evaluations explains, external reviews depend partly on whether evaluators can examine relevant model behavior and whether their findings influence release decisions.
California uses similar thresholds but different enforcement requirements
California's SB 53, enacted in September 2025, uses similar catastrophic-risk thresholds: death or serious injury to more than 50 people, or property damage or losses exceeding $1 billion from a single incident.
California's law primarily requires developers to publish safety frameworks and make specified disclosures, including incident reports. Anthropic's reported federal proposal would establish a separate mechanism for seeking judicial suspension of an advanced model. California also uses a “foreseeable and material” risk standard, so the numerical similarities do not establish identical requirements.
The approaches place different responsibilities on developers and regulators. California emphasizes documented safety practices and reporting obligations, while the proposed federal process would give Commerce a route to seek court intervention when an evaluation identifies an imminent catastrophic threat.
Anthropic has a commercial interest in federal AI safety rules
As a developer of advanced AI models, Anthropic has direct experience with the systems its proposed legislation would regulate. The company also has a financial interest in how federal safety requirements are defined. Mandatory external evaluations could impose costs that established AI developers are better positioned to absorb than smaller competitors.
An industry-wide evaluation requirement could also impose similar obligations on competing frontier AI developers. Depending on how Congress defines which models are covered, the legislation could affect development expenses and release schedules across the industry.
A federal suspension process could make enforcement procedures more predictable by identifying who can initiate a case and what evidence a court must examine. However, restricting intervention to a narrow set of circumstances could limit how regulators respond to serious risks below the catastrophic thresholds.
Federal legislation could also determine whether states retain authority to impose separate AI safety requirements. Anthropic's June framework recommends replacing state rules only in specified areas when federal protections are sufficiently comprehensive, while preserving the possibility of liability under state law. These recommendations do not establish which state requirements the reported bill would supersede.
Anthropic's proposal would create a judicial process for suspending AI models that present imminent catastrophic risks, while concentrating the authority to seek suspension within the Commerce Department. Whether this approach provides adequate protection depends partly on how quickly officials can respond and which measures remain available for threats below the proposed thresholds.
Congress would also need to resolve how independent evaluations are financed and how much authority individual agencies retain. As one of the companies subject to future frontier AI regulation, Anthropic has a direct interest in those decisions. The eventual legislation would determine both the government's ability to intervene and the compliance obligations imposed on AI developers.
