MASLEY: This is a situation where I really truly don't understand what it implies for AI safety.

NEYMAN: I think it really hinges on whether the Trump administration got serious about the risks or is executing on a political vendetta against Anthropic.

— Andy Masley and Eric Neyman, on the suspension of Fable 5 and Mythos 5, June 2026

I do not think it matters whether the administration is serious about AI risk or is executing a political vendetta against Anthropic. American AI policy is undergoing a national securitization. It is being pulled out of ordinary politics and into the part of the state that handles basic security and emergencies.

The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees. The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance.

Anthropic

I am less interested in the merits of this one restriction, which I think is a technical and political mistake, than in the direction of US AI policy. Which is something like an Aschenbrenner track:

This is the most important challenge for the national security of the United States since the invention of the atomic bomb. In one form or another, the national security state will get very heavily involved. [...] The Project will be the necessary, indeed the only plausible, response.

I am broadly in line with Aschenbrenner about where this is going, and not very relaxed about some of the costs.

The logic that pulls something into the national security state's orbit is also good at keeping it there, because the reason it went in rarely goes away. That's true even when security is used primarily as a political weapon, because it is not easy to walk back. It's going to be hard to redirect American AI policy off of this trajectory, though not impossible, and absent a concerted effort I expect it to be pulled fully into the security state's purview before the end of this administration.

None of that is to say a complete civilian reclamation would be desirable. There is a place for the national security state in making AI go well. Parts of the file, like the standards for securing model weights, should be led by the people who do security for a living, and frontier labs would concede as much. What I worry about is the rest of AI politics defaulting to them too.

The reason this matters so much is that we may be living through the most important century, which makes it matter enormously who manages it.

Think of any structure as a particular set of people and a particular way of reaching and enacting decisions. The national security state is a set of institutions built to manage security contests, first with the Germans and the Soviets, later with terrorist groups, and increasingly with China. That is real and important work, and they are good at it. But to do it the security state expects something close to subservience from those it deals with, and that is the wrong way to handle most of what advanced AI will bring up, from a new social contract to what it will mean to be human. They are not who I would want as a primary steward.

Even with the most obviously security-relevant issues the United States has been smart to keep the security services secondary, with civilian diplomats and the State Department leading on nuclear arms control and the services advising. A total security framing can flatten everything into country versus country, and many of the hardest problems here are not that at all, but about humans in relation to AIs.

The opacity of a totally securitized decision-making process is another problem. In the Fable and Mythos episode the administration and Anthropic have been working with very different expectations of each other. The same process will run through choices that matter far more, like whether to deploy a recursively self-improving system, how to handle a compute crunch, or whether to slow down. The failures there get larger and more spectacular. The national security state's process, behind classified channels and closed committees, would here shut out most of the people who founded and built the field of AI policy in the first place, some of whom argued for securitization themselves. Secrecy in the service's usual work is reasonable. Here it would be a bad feature.

Taken all the way, securitization also blocks the broad sharing of AI benefits. A partial one need not.

Advanced AI is going to remake warfare, command and control, and surveillance, very likely to decisive degrees, and you cannot treat a technology like that as sitting outside national security. To that extent the securitization is inevitable. But this can be thoughtfully done, you can reduce harms, and end up somewhere better than the default. A big part of doing that well is an empowered mediating body that understands the models and stays legible to the labs, something like the United Kingdom's AI Security Institute.

Security is always defined against an other, and to this rising power center the people who use, build, and study AI are suspect by default. It is sharpest for the foreign nationals who have done as much as anyone to push the frontier. I hope that being cast this way pushes more of them to think harder about the intermediate path to ASI, not to go quiet. I would like to see harder thinking about how the rest of the world comes to receive and use metered intelligence, and about how the country versus country contest should be managed toward the best outcome for the relationship between people and AI. For now there is a great deal of uncertainty, and not much seriousness. Will any of it even come up around Jay Clayton's nomination to lead American intelligence?