Anthropic Calls for Binding Audits as It Recasts AI as Strategic Infrastructure

Anthropic is shifting from transparency to enforcement

Anthropic has published a new policy argument that marks a notable change in posture: transparency requirements alone are no longer enough for frontier AI. In an essay titled Policy on the AI Exponential, CEO Dario Amodei argues that model capabilities are advancing fast enough to justify binding audits, formal disclosure requirements, and the power for a government agency to block systems that pose unacceptable risk.

The company paired the essay with two frameworks, one focused on regulating frontier AI and another on handling job losses that may accompany AI deployment. Anthropic says it is prepared to back the job-loss framework with significant funding. But the center of gravity in the new package is the regulatory turn. The company is no longer presenting frontier-model governance mainly as a matter of reporting and transparency. It is presenting it as a matter of technical inspection and pre-release control.

The case Anthropic is making

According to the source, Amodei’s argument begins with speed. He says AI capability growth is being driven by scaling laws and that within one to two years the world could see what he calls “Powerful AI,” described as “a country of geniuses in a data center.” That framing does two things at once. First, it emphasizes acceleration. Second, it elevates frontier AI from a commercial tool to something closer to strategic infrastructure.

Anthropic’s essay reportedly uses a Lord of the Rings analogy to explain the mismatch between rapid technical change and slow political response. In that story, Treebeard stands in for the political system, the warning voices are those pushing for early action, and the advancing threat is unregulated AI. The analogy is dramatic, but the policy implication is straightforward: if institutions move on ordinary timelines, they will react too late.

Why the company says its old approach no longer works

The source says Anthropic had previously emphasized transparency because the risks were not yet clear enough for more precise regulation. It supported transparency-focused legislation such as California’s SB 53, New York’s RAISE, and Illinois’s SB 315. That position has now changed.

As evidence, Amodei points to the experience with “Claude Mythos Preview,” which the source says disrupted the global cybersecurity landscape and demonstrated that frontier models can create real risks for the financial sector, critical infrastructure, and national security. Anthropic also expects serious biological and autonomy-related risks could emerge relatively soon. In other words, the company is arguing that theoretical concern has crossed into practical warning.

That is a meaningful threshold. A lab calling for stricter rules is not unusual. A lab saying its own recent model experience shows transparency is no longer sufficient carries more weight, even if critics will inevitably ask how much of the argument is principle and how much is strategy.

What Anthropic wants governments to do

The proposed framework centers on mandatory testing by qualified third parties across four risk categories: cybersecurity, biological weapons, loss of control over AI systems, and automated research and development that could accelerate those risks. Anthropic wants a government agency to have authority not only to receive disclosures, but to block or pull models that fail the standard.

The source says Amodei points to the Federal Aviation Administration as a model. The analogy is important. Aircraft are not approved for public use solely because the manufacturer says they are safe. They are inspected, tested, certified, and subject to intervention. Anthropic appears to be arguing that frontier AI should move toward a similar regime, where release is contingent on passing a technical risk review rather than merely satisfying notice requirements.

That is a higher-regulation vision than many tech companies have preferred in the past. It implies not only standards, but enforceable oversight and the possibility of shipment being delayed or denied. In policy terms, this is a shift from governance by transparency to governance by permission.

AI as a national security issue

The source says Anthropic increasingly depicts AI as a strategic weapon wielded by nation-states. That framing explains the urgency of the proposal. If frontier models are treated primarily as productivity tools, disclosure and market discipline might seem adequate. If they are treated as systems with implications for cyber offense, critical infrastructure, biosecurity, and geopolitical competition, then pre-deployment controls begin to look more plausible.

This is also why the four risk areas Anthropic highlights are so specific. The company is not calling for generic ethics review. It is calling for targeted testing where capability gains could spill into high-consequence misuse or systemic failure. The assumption is that the danger lies not in AI as an abstract category, but in a subset of frontier capabilities becoming strong enough to amplify already dangerous domains.

The politics of asking for stricter rules

Anthropic’s proposal will not end debate over who should write the rules or whether private labs are trying to shape regulation in ways that advantage incumbents. Those questions are unavoidable. But the new documents still matter because they show one of the best-known AI companies publicly concluding that its earlier regulatory preference has become insufficient.

That change is significant even if governments adopt only part of the agenda. Once major labs begin arguing for mandatory third-party audits and agency-level power to block releases, the debate moves beyond voluntary commitments. It becomes a question of institutional design: who tests, who decides, what triggers intervention, and how much capability evidence companies must disclose before launch.

A policy pivot with broad implications

The most consequential part of Anthropic’s intervention may be the timing. The company says the next one to two years could bring systems much more capable than those currently in public use. If policymakers accept that timeline, the window for building oversight structures is short. If they reject it, Anthropic’s framework may still influence the shape of future rules by normalizing the idea that frontier models need pre-release audits.

Either way, the company is helping redefine the center of the AI policy conversation. The argument is no longer simply that AI firms should be more open. It is that some models may be too consequential to release without external technical review and state authority to say no.

That is a far more assertive theory of governance than the one that dominated many early AI policy debates. It reflects both the rising ambition of frontier systems and the growing belief among some of their builders that transparency by itself cannot contain what comes next.

This article is based on reporting by The Decoder. Read the original article.