Anthropic Pitches AI Safety Bill to Congress

10 Min Read
Anthropic branding at a technology event as the company pitches an AI safety bill to Congress
Anthropic is urging lawmakers to adopt enforceable safeguards for the most powerful AI models. Source: American News Brief media library; original credit Carlos Barria/Reuters.

Anthropic is quietly pitching U.S. lawmakers on a federal AI safety bill that would give the government limited authority to block the release or use of the most powerful artificial-intelligence models when narrowly defined catastrophic risks are imminent. The proposal would move the national AI debate from voluntary corporate commitments toward enforceable rules while attempting to avoid a broad licensing regime for the entire industry.

The draft would allow the government to intervene when a frontier model poses a foreseeable and imminent risk of more than $1 billion in damages or of killing or seriously injuring more than 50 people, according to reporting on the proposal. The qualifying risk would need to involve areas such as biological weapons, a model evading its developer’s controls or a major cyberattack.

Anthropic AI Safety Bill Sets a High Intervention Threshold

The proposal is notable for how high it sets the bar for federal intervention. Rather than giving regulators broad authority to approve every major AI model before release, the bill would reserve blocking power for cases involving extraordinary and imminent harm.

That structure appears designed to answer two competing political concerns at once. AI safety advocates want the government to have real authority when frontier systems become dangerous, while opponents of heavy regulation warn that a general licensing system could slow innovation, entrench the largest companies and weaken the United States against China.

The bill would also require developers of frontier systems to conduct testing, disclose relevant findings and submit to independent evaluations. Outside evaluators would register with the Commerce Department, creating a formal role for third-party assessment rather than leaving safety judgments entirely to the companies building the models.

Anthropic Pushes Regulation After Trump Resisted New Rules

The proposal puts Anthropic directly into a policy debate the White House has approached cautiously. American News Brief previously reported that Trump rejected calls for a broad new AI regulatory push, arguing that excessive restrictions could slow American companies while Chinese competitors continue advancing.

Anthropic CEO Dario Amodei speaks as debate grows over federal AI safety rules
Anthropic CEO Dario Amodei has pushed for stronger safeguards and independent oversight for frontier AI systems. Source: American News Brief media library; original credit Yves Herman/Reuters.

That disagreement has created an unusual dynamic in Washington. Some of the companies developing the most capable AI systems are asking for stronger federal oversight, while the administration has emphasized voluntary safeguards, competition and national technological leadership.

Anthropic’s draft tries to narrow that gap by focusing regulation on frontier models and catastrophic-risk thresholds rather than ordinary software. The company is not asking Congress to regulate every chatbot, recommendation system or business automation product under the same framework.

Independent Evaluations Would Become Part of the System

One of the most consequential elements is the proposed role for independent evaluators. Frontier-model developers would be expected to test high-risk capabilities and permit qualified outside reviewers to assess whether safeguards work as claimed.

That matters because many of the hardest AI safety questions cannot be answered by reading a company’s public documentation. Evaluators may need access to models, internal testing environments and technical controls to determine whether a system can bypass restrictions, assist with dangerous biological work or conduct sophisticated cyber operations.

Anthropic has already been moving toward deeper external review. The company recently granted the nonprofit Model Evaluation and Threat Research unusually broad access to some internal systems, an arrangement intended to test whether independent oversight can coexist with protection of customer data and proprietary technology.

Biological and Cyber Risks Drive the Proposal

The draft focuses on risks that can translate into direct physical or economic harm. Biological misuse, cyberattacks and systems that escape or evade developer controls are among the scenarios that would justify intervention under the proposed framework.

Anthropic has been especially vocal about the possibility that advanced models could lower technical barriers in biology. American News Brief recently covered Anthropic’s expansion into physical life-sciences research, a move that highlights both the scientific value of advanced AI and the safety questions that arise when models gain deeper knowledge of laboratory workflows.

The same dual-use problem exists in cybersecurity. A system capable of finding software vulnerabilities can help defenders patch systems faster, but similar capabilities can also be used to automate attacks or accelerate exploitation if controls fail.

Congress Still Lacks a Comprehensive AI Framework

Congress has debated artificial-intelligence regulation for years without creating a comprehensive federal regime. Lawmakers have advanced narrower bills on deepfakes, children’s safety, national security, export controls and government use, but the broader question of how to regulate frontier models remains unresolved.

That regulatory vacuum has pushed companies and states to experiment with their own standards. Supporters of federal legislation argue that a national framework could reduce uncertainty and prevent a patchwork of conflicting state rules, while critics warn that federal preemption could wipe out stronger local protections.

The Anthropic AI safety bill would add another option to that debate by focusing the federal role on the most capable systems and the most severe risks. Whether lawmakers view that as appropriately narrow or too deferential to industry will likely determine how much support it attracts.

Big Tech Regulation Raises Competition Questions

Any proposal written or promoted by a major AI developer will face scrutiny over regulatory capture. Large firms can absorb compliance costs, hire specialized legal teams and build elaborate testing programs more easily than startups, so rules that appear safety-focused can also create competitive barriers.

Anthropic’s challenge is to persuade lawmakers that the framework targets genuine external risks rather than protecting incumbent firms. High intervention thresholds and a focus on frontier systems may help that case, but details about which companies qualify and how evaluation costs are allocated will matter.

Competition policy is especially important because the U.S. government is simultaneously trying to preserve American leadership against China. Regulations that are too weak could leave the country exposed to preventable failures, while rules that are too costly or rigid could push research and investment elsewhere.

The White House and Congress Face a Strategic Choice

The proposal arrives as AI becomes more central to national security, economic policy and industrial strategy. Frontier systems are increasingly relevant to coding, scientific research, intelligence analysis and cyber defense, raising the cost of both under-regulation and over-regulation.

Trump’s administration has generally preferred voluntary commitments and targeted government action instead of a broad statutory regime. Anthropic is now arguing that some risks require enforceable authority before a catastrophic event occurs, even if that authority is reserved for exceptional cases.

The debate is therefore shifting from whether AI should be regulated at all to what kind of regulation can preserve innovation while creating meaningful accountability. A narrowly tailored federal system could appeal to lawmakers who reject both a hands-off approach and a sweeping licensing bureaucracy.

What Happens Next

The draft is not yet federal law, and Congress is approaching the midterm elections with limited time for major legislation. Lawmakers could use the proposal as a starting point for hearings, merge elements into existing bills or leave the issue for the next Congress.

The political landscape may also change after Nov. 3. Control of the House or Senate could determine which committees lead the debate, how aggressively Congress challenges the White House and whether industry-backed safety proposals gain traction.

For Anthropic, the immediate objective appears to be establishing a concrete legislative model before a crisis forces Washington to act under pressure. The Anthropic AI safety bill is therefore less important as a finished product than as a test of whether Congress can agree on a limited federal role for the most powerful AI systems before their capabilities advance further.

Share This Article