Anthropic's bill for Congress: a regulator with its hands tied and a brake that takes 72 hours

By Crowdfund Insider (citing Politico and Mashable) · October 18, 2026.

Anthropic has been privately circulating a draft federal bill for frontier AI systems among lawmakers on Capitol Hill. According to Politico, which obtained the confidential document and spoke with three people familiar with the discussions, the text closely tracks the Advanced AI Framework, the internal framework the company has been pitching to federal and state authorities since mid-2025. An Anthropic spokesperson described it as reflecting "sensible policies" that Congress should pass.

The core of the draft is the intervention threshold. According to the reporting, the government could only block a model's launch or continued use in the face of a foreseeable and imminent risk of more than $1 billion in damages or of death or serious injury to more than 50 people. In addition, the harm would have to come through one of three pathways: the creation or release of a biological weapon, the model escaping or evading its developer's controls, or a cyberattack.

The enforcement mechanism is also narrow. Independent evaluators, registered with the Department of Commerce, would conduct periodic reviews of the labs and their most advanced models. Developers could pay them, provided the fee does not depend on the outcome. If evaluators detect an "imminent enumerated risk," they must notify Commerce within 72 hours, and Commerce could ask a federal judge in the District of Columbia to suspend the model. The text would bar any other federal department, agency or official from halting the development, deployment or internal use of a model.

The draft would also include preemption of certain state requirements on risk disclosure, third-party evaluation and critical incident reporting. In exchange, it would require safety frameworks, transparency reports, whistleblower protections and incident reporting to Commerce. According to the outlet, some of the language echoes California's SB 53, which Anthropic supported.

The legislative landscape is already crowded. There are bipartisan initiatives in both chambers, such as the FRONTIER Act in the House of Representatives (backed by Lori Trahan and Jay Obernolte) and negotiations in the Senate involving John Thune, Amy Klobuchar and Ted Cruz. According to the source, some safety advocates and lawmakers want a broader government role in reviewing models.

AI governance researcher Mana Onishi raises the central question: if an advanced system began to display dangerous autonomous behavior, would there be clear authority and the technical means to stop it in real time? The narrow triggers, limited harm pathways and small number of actors able to seek court relief leave the question of response speed open.

**Our take.** A company proposing the regulation it wants to comply with is neither new nor, in itself, illegitimate. What's more, a concrete text on the table, with mandatory transparency, whistleblower protections and external evaluators, is better than a vacuum. But it should be read for what it is: a proposal from an interested party. It cannot be claimed that it seeks to capture the regulator; what we can point out is that its design concentrates the power to halt a model in a single channel and subjects it to a very high bar.

The asymmetry lies in the numbers. A threshold of $1 billion and 50 deaths protects against disaster, but leaves out serious and more diffuse harms, such as mass fraud, manipulation or infrastructure failures that don't fit the three pathways. And the 72-hour window plus a court process is a cadence designed for litigation, not for an incident that unfolds in minutes. That is Onishi's objection, and we think it is the most serious one.

Broadly, AI governance advances through three channels: ex ante regulation, ex post litigation and audited disclosure. This draft bets on the third, along with federal preemption that simplifies compliance, something valuable for anyone operating in 50 states. The risk is that this simplification is bought by watering down stricter state requirements.

Our position is consistent with our long-standing line: regulate real capability, based on evidence, not panic. That requires genuine technical instruments: independent evaluations, demonstrated capacity to intervene and a procedure that runs before procurement or deployment, not only after a catastrophe. In the short term the debate will be rough and the interests obvious. In the long term, if AI brings us closer to curing diseases and to abundance, having credible rules that maintain public trust will be part of how we get there. What to watch is how much of this draft survives once it is weighed against the proposals of Trahan, Obernolte, Thune, Klobuchar and Cruz.