The Mythos Fight Shows a New Kind of AI Concentration Risk

The recent fight over Anthropic’s Mythos frontier model may look like a one-off Washington kerfuffle. It is probably better understood as the latest step in a much longer process.

For at least four years, the United States has been trying to shape the AI race with China through national security tools. The first stage was hardware: restrict China’s access to the most advanced chips and chipmaking equipment. The Commerce Department’s 2022 controls on advanced computing chips and semiconductor manufacturing items were the clearest early signal.

The second stage was visibility. President Biden’s 2023 AI executive order created a category of “dual-use foundation models” and required companies building the most powerful systems to provide the federal government with more information about training, model security, and safety testing. That moved the policy focus from the inputs to AI toward the models themselves. The question was no longer only who could get the chips. It was also what the frontier labs were building, how powerful those systems were becoming, and how securely they were being controlled.

The third stage was partnership. Frontier AI companies have increasingly been pulled into defense, intelligence, cybersecurity, and critical infrastructure work. Anthropic’s Claude Gov models for national security customers are one example. The larger pattern is that the federal government is not just regulating the leading AI companies. It is also becoming a major customer, partner, evaluator, and operational user.

Now we appear to be entering a fourth stage: model access itself.

That is what makes the Mythos episode important. According to Reuters, the U.S. government ordered Anthropic to suspend access to its Fable 5 and Mythos 5 models over national security concerns. A later Reuters report said the government allowed Anthropic to redeploy Mythos to selected “trusted” U.S. organizations. That is a different kind of governance move. It is not just safety testing. It is not just export control on chips. It is the government reaching more directly into who can use a frontier model, under what conditions, and for what purposes.

Some of this may be necessary. Cybersecurity is genuinely dual use. A model that helps defenders find vulnerabilities faster can also help attackers do the same. And if frontier AI is becoming part of military, intelligence, and critical infrastructure systems, it is not realistic to pretend national security will stay outside the room.

But this is also exactly where democratic vigilance matters most.

National security is one of the strongest justifications government has for secrecy and control. Sometimes secrecy is justified. But if national security becomes the main pathway for governing frontier AI, then some of the most important decisions about the technology may happen out of public view: which models are allowed to be released, which users are trusted, which companies get early access, which foreign users are excluded, and which private labs become embedded in federal security operations.

That would be a major concentration signal.

The usual AI concentration story focuses on big companies: who owns the models, the data centers, the chips, the cloud platforms, and the distribution channels. That story is still important. But Mythos points to another concentration path. Power may also concentrate through the national security state, especially if government and a handful of frontier labs become tightly interdependent.

The ultimate version would be a federal order placing frontier models fully under government control. We are not there. But the direction of travel is worth watching. The White House’s June 2026 order on advanced AI innovation and security says it does not authorize mandatory licensing or preclearance for model release. At the same time, it calls for trusted partners, early access, and close coordination with frontier model developers. That is a narrow line to walk.

The harder question is not whether national security concerns are real. They are. China’s rapid progress in cyber-capable AI, including reports that Z.ai’s GLM-5.2 has matched Mythos in some cybersecurity tasks, makes the concern more concrete, not less. But that only raises the stakes for public rules. If national security is going to shape who gets access to frontier AI, then the public needs at least some visibility into the broad rules of the road.

When can the government restrict a model? Who decides? What standards are used? How long do restrictions last? What rights do affected users or companies have? What kind of congressional or independent oversight exists? And how do we keep necessary security controls from becoming a permanent, opaque control layer over the most important technology of the next decade?

The Mythos controversy is not just about one model. It is a warning that AI governance may increasingly be built through national security exceptions. If that happens quietly, the public may not notice the control layer being constructed until it is already in place.