Anthropic’s Regulatory Capture Strategy Is Clever, Cynical, and Not in the Public Interest
Anthropic’s regulatory capture strategy is clever, cynical, and not in the public interest.
In mid-June 2026, the U.S. government delivered an unusually direct demonstration of why.
Days after Anthropic released its latest flagship models, Fable 5 and Mythos 5, the Commerce Department issued an export control directive barring all foreign nationals, anywhere in the world, including inside the United States and even within Anthropic itself, from accessing them. Unable to easily filter users by nationality across its platforms, Anthropic was forced to abruptly disable the models for every customer worldwide.
This was not the careful, surgical regulation Anthropic had spent years advocating. It was the correction of a grizzly bear: impolite, heavy-handed, and impossible to forget.
Anthropic’s strongest argument is not mere fearmongering. It is the claim that certain catastrophic risks, especially those involving biological weapons or advanced autonomous cyber capabilities, cannot be undone once the weights of a sufficiently powerful model are released or leaked. In this view, the only responsible path is to keep frontier systems tightly controlled, with rigorous evaluations and regulatory guardrails that limit who can develop and deploy them.
The trouble with this position is that the solution it proposes: concentrating the most advanced capabilities inside a few closed labs while building regulatory barriers that only those labs can afford, and makes the very risks it fears more dangerous, not less. Closed labs become single points of failure. When frontier models live behind high walls and heavy compliance requirements, a single breach, insider action, or state compromise carries outsized, potentially irreversible consequences. At the same time, the broad, independent scrutiny that open development naturally produces is throttled. Multiple teams studying the same problems, forking models, and iterating quickly tends to surface issues faster than any single authorized safety organization can. Distribution itself becomes a form of resilience. Anthropic’s preferred regime does the opposite: it narrows the number of actors who can see, test, and respond while raising the stakes of any failure.
This is why the regulatory push functions as capture. By championing safety frameworks that impose large fixed costs inlcuding mandatory evaluations, scaling policies, extensive documentation, and approval processes. Anthropic seeks to create rules that smaller players and open-source efforts cannot meet. The practical result is a market in which real innovation is largely restricted to the same handful of companies that can staff large legal and compliance teams. That outcome serves the incumbents far more reliably than it serves public safety.
The June 2026 government action reveals a deeper problem with this entire strategy. An advocate for formal regulation might argue that the episode simply shows the need for clearer, more predictable rules instead of blunt national security tools. That misses the core issue. Once companies successfully normalize the idea that the state should exercise wide discretionary power over frontier AI development, whether through export controls, mandatory evaluations, or deployment restrictions, they surrender meaningful control over how that power will ultimately be used. Administrations change. Priorities shift. The same tools and precedents that one set of labs helped legitimize can be turned against those labs, against their competitors, or against open development when political or bureaucratic incentives move. There is no reliable mechanism to keep the state’s hand steady or narrowly calibrated once that authority has been invited into the room.
The public interest lies in accelerating the diffusion of capability rather than concentrating it behind regulatory moats. That means supporting open weights and open inference infrastructure that allow the widest possible community to examine, improve, and defend against risks. It means focusing regulatory attention on concrete high-stakes misuse rather than broad capability thresholds that only the largest players can clear. And it means recognizing that resilience against serious harms is better achieved through many independent actors and rapid iteration than through a small number of authorized stewards operating under rules they helped design.
Anthropic’s regulatory capture strategy is clever, cynical, and not in the public interest. We should reject it.
Pages
