Anthropic shares draft AI bill with high thresholds for federal action

The proposed safety thresholds

Anthropic, the company behind the Claude AI models, recently shared a confidential draft bill with US lawmakers regarding frontier AI safety. This proposal serves as the company's own blueprint for how the federal government should police the development of advanced systems. Rather than broad oversight, the bill focuses on specific triggers that would allow regulators to step in.

Under the draft, the federal government would only be permitted to block the release or use of a model if it posed a foreseeable and imminent risk. Specifically, the proposal sets a high bar requiring more than $1 billion in damages or a threat to kill or seriously injure more than 50 people. These requirements mean that risks which are plausible but not currently looming would fall outside the reach of the bill.

The scope of these interventions is also strictly limited to three specific scenarios. The bill only covers models that could help create or release biological weapons, models that slip out of a developer's control, or those involved in cyberattacks. By narrowing the types of dangers covered, the draft creates a very specific path for federal intervention rather than a general safety standard.

Governance and reporting structures

The draft identifies the Department of Commerce as the sole agency with the authority to go to court to pause a specific model's deployment. This concentration of power means other federal agencies would lack the ability to intervene in these specific cases. This design aims to streamline the process, though it creates a single point of failure if the agency is under-resourced or slow to act.

To ensure these risks are identified, the proposal includes a requirement for repeated independent evaluations of frontier models. If an evaluator discovers an imminent risk, they would have a 72-hour window to report those findings to the Department of Commerce. This structure is intended to create a formal pipeline for safety concerns to reach the government quickly.

Additionally, the bill proposes a framework for incident reporting and protections for whistleblowers. These elements are designed to encourage transparency within the companies developing these models. By formalizing how problems are reported, the bill seeks to create a predictable environment for both developers and regulators.

Origins and industry lobbying

This proposal is not a sudden move by Anthropic but rather the latest step in a months-long lobbying effort. The company has been promoting its Advanced AI Framework to lawmakers since June 2026. The draft also draws heavily from California’s SB 53, a law signed in 2025 that established similar thresholds for AI safety.

Critics have pointed out that the bill exists largely to address concerns raised by Anthropic’s own staff and leadership. Some former employees and current executives have warned that the systems they are building could pose catastrophic risks to humanity. The draft is seen by some as a transparent attempt to calm those public fears by providing a controlled regulatory path.

The debate over this bill highlights a recurring question in Washington regarding who should write the rules for AI. Critics argue that governments should be the ones to set the standards rather than letting the companies themselves draft the legislation. This is particularly notable because Anthropic is one of the loudest voices warning about the dangers of the technology it helps create.

The path forward for regulation

The draft bill currently exists as a pitch to lawmakers and has not been formally introduced as legislation. It will now compete with other proposals, such as the bipartisan FRONTIER Act, which offers different oversight mechanisms. The contrast between Anthropic's high-bar thresholds and these other proposals will likely shape the final federal standard.

Anthropic has acknowledged that certain parts of the draft language still need work and has committed to refining the wording by October 18, 2026. Whether the company lowers the $1 billion threshold or broadens the types of risks covered will be a key signal of its response to the current backlash. Any changes to these numbers would directly impact how much power the government has to intervene.

For now, the draft remains a statement of what one AI company wants the rules to be. The next major milestone will be whether any member of Congress formally introduces the bill into the legislative process. Until that happens, the proposal serves as a primary example of the industry's preferred approach to safety and oversight.