Anthropic Draft Bill Limits Federal AI Shutdown Authority

The tension between executive agency discretion and statutory precision is a recurring pattern in technology governance. Historically, emerging industries have sought to replace broad administrative authority with specific statutory thresholds to reduce regulatory uncertainty, often following periods of aggressive enforcement or export control actions. This dynamic mirrors precedents in telecommunications and financial services, where industry-drafted frameworks were introduced to Congress after firms faced unpredictable licensing decisions or operational restrictions. The strategy typically involves proposing high quantitative triggers for intervention to ensure that only extreme scenarios warrant government disruption of business operations. Conversely, legislative counter-proposals often emphasize retaining flexible agency authority to address novel harms that do not fit predefined categories. This cycle reflects a structural negotiation over the allocation of risk: industry prioritizes predictability and capital deployment, while legislators balance innovation incentives against the political cost of under-regulation. The current dispute fits this established pattern of post-enforcement regulatory codification.

On October 18, 2026, Politico published details of a confidential draft bill authored by Anthropic that seeks to codify strict limitations on federal authority to suspend or shut down frontier artificial intelligence models. According to the leaked document, the proposed legislation would permit government intervention only when a model presents a foreseeable and imminent threat meeting specific quantitative thresholds: either more than $1 billion in damages or the death or serious injury of over 50 people. Furthermore, the draft restricts qualifying catastrophic risks to three narrow categories: the creation or release of biological weapons, model escape from developer controls, or cyberattacks. Under this framework, federal authorities would be required to obtain judicial approval before suspending a model, and independent evaluators would be tasked with overseeing compliance rather than direct regulatory enforcement.

Anthropic has positioned this draft as a legislative embodiment of its Advanced AI Framework, a policy initiative the company began promoting to federal and state lawmakers in June 2026. An Anthropic spokesperson stated that the proposal represents sensible policies Congress should adopt, arguing it provides necessary legal clarity for developers while preventing arbitrary government interference. The company’s push for defined guardrails follows reported incidents in August 2025 and July 2026 where U.S. export controls allegedly sidelined two advanced Anthropic models, intensifying industry concerns regarding executive overreach and regulatory unpredictability.

The leak has triggered significant opposition from lawmakers who argue the thresholds are too permissive. Senator Maria Cantwell has advocated for stronger federal review processes than those outlined in the industry-drafted proposal, suggesting the current draft insufficiently protects public safety. Senator Bernie Sanders has taken a more maximalist stance, calling for a pause on advanced AI development entirely rather than establishing narrow intervention thresholds that he views as inadequate for managing existential risk. These critics contend that tying shutdown authority to such high damage or casualty benchmarks effectively immunizes developers from accountability for harms that fall below these catastrophic lines but remain socially destructive.

Simultaneously, competing legislative efforts highlight the fragmentation of the current regulatory landscape. Representative Lori Trahan has advanced the FRONTIER Act, a separate bill that authorizes broader government intervention and includes provisions to override certain state laws, directly contrasting with Anthropic’s more restrictive approach. In the Senate, John Thune is working with bipartisan colleagues on separate legislation targeting catastrophic AI risks, indicating that Congress is actively considering multiple frameworks. Root Nation reported that Anthropic’s proposal also includes provisions to limit the effect of some state laws, suggesting an intent to establish federal preemption in exchange for accepting the specified catastrophic risk thresholds.

The controversy centers on whether private industry should define the legal parameters of its own regulation. While Anthropic argues that clear, high-threshold rules prevent capricious enforcement and encourage innovation, critics view the draft as an attempt to legislatively entrench industry preferences and raise the bar for regulatory action beyond what public safety requires. The requirement for judicial approval adds another layer of procedural friction to government oversight, which defenders see as due process and opponents view as a delay mechanism during emergent crises. As of late October 2026, the draft remains a proposal without a formal sponsor, but its leakage has crystallized the debate over how much discretion federal regulators should retain versus how much certainty industry demands.

Anthropic and frontier developers stand to gain predictable operating conditions and protection from discretionary export controls if the $1B/50-casualty thresholds are adopted. Federal regulators risk losing preemptive intervention authority for harms below these high bars. Society faces potential exposure to systemic risks that are severe but do not meet the defined catastrophic criteria. The competing FRONTIER Act and Senate efforts suggest legislative outcomes remain fluid, but Anthropic's draft sets a concrete anchor point for negotiations. If enacted as drafted, the bill would fundamentally reallocate risk tolerance from the public sector to private developers, making government response reactive rather than preventive for all non-existential threats.