Anthropic Pitches Draft AI Safety Bill to Congress Mandating High-Bar Vetoes

The boundary between unregulated artificial intelligence and federal intervention will ultimately be drawn by the specific damage thresholds written into US law. To ensure it has a hand in defining that boundary, AI developer Anthropic has drafted its own legislative framework and quietly circulated it among lawmakers on Capitol Hill.[1][2]

The confidential draft, first reported by Politico on October 16, 2026, represents a major frontier AI lab attempting to establish the first federal framework for AI safety. It proposes a narrowly defined "off-switch" that would grant the US government the authority to block dangerous deployments.[1][3]

However, the bill sets a deliberately high bar for federal action. Under the proposed legislation, the government could only halt a model's release if it poses a foreseeable and imminent risk of causing more than $1 billion in damages, or of killing or seriously injuring more than 50 people.[1][4]

The legislation strictly limits the types of threats that would qualify for this intervention. The risks must stem directly from the creation of biological weapons, a cyberattack, or an AI system actively evading its developer's controls, according to the circulated draft.[1][5]

Establishing an oversight mechanism

To monitor these risks, the bill mandates that independent evaluators register with the US Commerce Department. These auditors would be tasked with periodically testing frontier models and the laboratories that develop them, ensuring compliance with the established safety thresholds.[1][3]

If an evaluator identifies an "imminent enumerated risk" during their assessment, the legislation requires them to report the finding to the Commerce Department within 72 hours. The labs are permitted to pay these evaluators, provided the compensation is not tied to the audit's outcome.[3]

The enforcement mechanism is highly centralized, stripping authority from broader federal bureaucracies. Only a federal judge in the District Court for the District of Columbia would have the power to order a model's suspension, and only following a formal request from the Commerce Department.[1][3]

"The language explicitly prohibits any other US government department, agency, or officer from suspending development, deployment, or internal use," Reuters reported, highlighting the bill's intent to prevent fragmented regulatory actions across different federal bodies.[3]

Preempting state-level regulation

A critical component of Anthropic's proposal is a federal preemption clause designed to override state-level AI safety laws. This provision would limit the ability of individual states to impose their own testing, disclosure, and safety requirements on AI developers.[1]

The push for federal preemption arrives as state legislatures, such as those in Pennsylvania and California, advance their own AI safety bills toward committee hearings. By establishing a single national standard, Anthropic aims to avoid a patchwork of conflicting state regulations.[1]

The legal principle of federal preemption allows national law to supersede state law in areas where Congress has chosen to regulate. The exact extent of these restrictions on state authority will depend on the final language adopted by lawmakers.[1]

Anthropic's lobbying effort coincides with broader congressional attempts to govern the rapidly advancing technology. In late September 2026, a bipartisan group of senators introduced the Artificial Intelligence Risk Management and Security Act, which proposed a federal AI Safety Board.[1]

Shaping the regulatory landscape

The draft bill essentially translates Anthropic's internal Advanced AI Framework into proposed federal law. An Anthropic spokesperson confirmed that the circulated legislation reflects the safety and scaling proposals the company released earlier in the year.[1]

By setting the intervention threshold at $1 billion in damages or 50 casualties, the company is attempting to balance innovation with catastrophic risk prevention. The draft focuses entirely on extreme, frontier-level threats rather than near-term issues like algorithmic bias or copyright infringement.[1][2]

The privately circulated draft has not yet been formally introduced in either chamber of Congress. For the framework to advance, lawmakers must agree to sponsor the measure, subjecting it to standard congressional consideration and debate.[1][4]

If sponsored, the legislation would require approval from both the House of Representatives and the Senate, followed by the president's signature. The proposal is widely viewed as an opening bid by Anthropic to anchor the debate before Congress drafts its own comprehensive rules.[1][5]

Industry and political reactions

The move highlights a growing trend of technology companies actively drafting the legislation meant to regulate them. While Anthropic positions the bill as a necessary safeguard against catastrophic risks, critics argue it allows the industry to dictate its own oversight terms.[1][2]

Some observers have noted that the bill's high threshold for intervention means it does not function as a general-purpose "kill switch" for AI. Instead, it creates a narrow legal pathway that protects developers from arbitrary government shutdowns while addressing extreme scenarios.[1][3]

The debate over AI regulation has intensified following public warnings from researchers about the existential risks posed by advanced models. Lawmakers are increasingly pressured to establish a regulatory framework before the technology outpaces the government's ability to monitor it.[1][4]

Ultimately, the success of Anthropic's legislative push will depend on how lawmakers weigh the proposed thresholds against competing regulatory models. As the 2026 legislative session progresses, the draft will likely serve as a baseline for upcoming negotiations on frontier AI safety.[1][5]

Formalizing the evaluation process

The proposed framework also addresses the mechanics of transparency without mandating public disclosure of proprietary algorithms. Under the draft, frontier AI developers would be required to test their models and disclose findings to the registered evaluators, rather than open-sourcing their core technology.[1][3]

This approach aligns with the national security concerns surrounding advanced AI systems. Earlier in 2026, the US Commerce Department utilized export controls to restrict foreign access to certain frontier models, demonstrating the government's willingness to act unilaterally when risks are perceived.[1]

By formalizing the evaluation process and defining the exact triggers for intervention, Anthropic's bill seeks to replace ad-hoc executive actions with predictable statutory law. Developers would have a clear understanding of the safety benchmarks required to deploy their systems domestically.[1][2]

The coming months will reveal whether Congress adopts Anthropic's high-bar approach or pursues more stringent, preemptive licensing regimes. The outcome will dictate the operational environment for the entire US artificial intelligence sector for the foreseeable future.[1][4]