Hugo Lowell, Lily Hay Newman, and Maxwell Zeff, reporting for Wired (Apple News+):

Trump administration officials concluded talks with Anthropic on Monday without lifting export controls that were imposed last week on the company’s most advanced AI models in response to jailbreaking concerns, according to three people briefed on the matter.

The administration continues to believe that there are ways to disable some of the guardrails on Anthropic’s Claude Fable 5, effectively allowing users to access the more powerful cybersecurity capabilities of the company’s Mythos model, the people said.

Anthropic has said for days that the administration’s concerns are overblown, a position it reiterated in working group meetings held at the Commerce Department with government researchers from the Center for AI Standards and Innovation and the Office of the National Cyber Director, Sean Cairncross, one of the people said.

There are a few angles to this conundrum:

  1. Why did the Commerce Department suspend Claude Fable 5?
  2. Was Anthropic’s “fearmongering” responsible?
  3. Does the Trump administration have the legal right to impose export controls in this way?

First, why? The Trump administration’s official word is that the guardrails on Claude Fable 5 — which prevent users from using the model for certain biological and cybersecurity problems — can be easily jailbroken. Amazon, one of Anthropic’s key inference and investment partners, discovered the guardrails and, instead of reporting them to Anthropic directly, went straight to the Commerce Department. This is uneducated speculation, but that sounds like intentional sabotage to me. Either way, the Trump administration, after hearing about the jailbreak, ordered Anthropic to suspend Claude Fable 5 for all non-U.S. nationals, including those legally in the United States.

This reaction is nonsensical: For one, it is virtually impossible to enforce a domestic export control like this. (I’ll elaborate on this in a moment.) Second, Americans are absolutely capable of sabotaging U.S. military action and compromising national security. What stops a foreign adversary from bribing an American who has access to Claude Fable 5? So the rationale for this is clearly not national security or genuine concern, and fortunately, the Trump administration isn’t tight-lipped on culture-war issues. Defense Secretary Pete Hegseth posted this incredible message on X right after the export controls were hastily announced:

Three months ago, @DeptofWar kicked @AnthropicAI out of our building—forever.

Every passing day proves why that was the right move.

The U.S. government is waging a war against an American company. Media outlets have been caught up in the technical detail of the export controls — and, to an extent, so have Anthropic representatives negotiating — because they have “Silicon Valley brain.” Washington journalists know what’s happening here: the Trump administration is exacting revenge after Anthropic refused to let the Defense Department use Claude models for “all lawful purposes.” There is zero ambiguity here for why the Commerce Department decided to levy such a punitive, impossible punishment. This is not a matter of national security — this is the defense secretary laying it all out in the open.

So I think that makes the second angle moot, but only to an extent: Was Anthropic’s “fearmongering” about Claude Mythos responsible? I think yes, but not in the most direct way. Regardless of whether Anthropic said Claude Mythos and Claude Fable 5 were dangerous or not, they’re the most capable large language models ever created, in every benchmark. Third-party security researchers would inevitably try to jailbreak them because of their capability, not Anthropic’s advertising. But Anthropic has also advocated heavily for government oversight under the Trump administration, with Dario Amodei, Anthropic’s chief executive, saying the administration should have the power to cancel the deployment of its models if it deemed such action necessary. This is the only argument from the pro-Trump administration camp I find convincing: Anthropic wanted this. From Amodei’s essay titled “Policy on the AI Exponential”:

Frontier AI models, like airplanes, should be required to go through technical testing and auditing, and their release should be blocked or reversed as a threat to public safety if they do not meet high standards of safety. I am grateful to see the Trump administration’s Executive Order move incrementally towards a greater role for government in AI, though Anthropic’s proposal recommends even further action.

There are really two ways to approach artificial intelligence regulation: the OpenAI approach and the Anthropic one. OpenAI has no morals, ethics, or responsibilities. Its only goal is to make as much money as possible and turn the Great Plains into one gigantic data center. If the U.S. military wants unfettered access to OpenAI models, OpenAI will happily provide it. If the Trump administration wants to tune out “woke” information in the models’ weights, OpenAI will oblige gracefully. OpenAI will do anything to score an extra government contract and inch one step closer to world domination. It has always been that way.

Historically — before this passage in Amodei’s essay — the Anthropic approach has been quite meritocratic. It welcomed the government’s thoughts — particularly around protecting users’ economic and civil rights — but ultimately, it felt the final decisions of how those models were made and distributed should be in the hands of experts, not the government. Anthropic would rather lose government contracts worth billions of dollars than have its models make autonomous decisions in war — that was its red line and it refused to cross it. But after the kerfuffle with the Defense Department in February, I feel Anthropic emerged unnerved. It learned what happens to companies that oppose the administration. It silently abandoned this staunch authoritarian stance and loosened its grip on its models. So now, we’re here, where Anthropic is split between controlling its own models and letting the government regulate them. Anthropic had a good thing going and allowed the U.S. government to destroy it.

So, the third angle: Is any of this even legal? I am not a lawyer, but I err toward no. The government appears to be relying on the Export Administration Regulations, a set of laws that govern export controls. The EAR affects products that have dual military and civilian uses, such as LLMs. The government then imposes an export restriction on a usually narrow scope of countries, like China or North Korea. These restrictions may also be applied to foreign nationals from those countries. The Commerce Department has instead used the EAR to issue a blanket ban on all foreign use of Claude Fable 5, including many of Anthropic’s own employees who worked on the model. This is an outlandish interpretation of this law.

It hasn’t been well-conceived, either. Do dual citizens count? And how does an export restriction in the first place ensure foreign nationals don’t use the model? How does a company enforce such a restriction domestically? The most likely timeline is that the government wanted to punish Anthropic by kneecapping its flagship model, realized it didn’t have a law that allowed it to — thanks to the First Amendment — and used export controls inappropriately to achieve the same effect. I reckon that if Anthropic sued the government, it would win at least a temporary injunction. (Again, I am not a lawyer.) Banning people legally in the United States from using a private company’s technology is truly unprecedented. This whole thing is unprecedented — but it might just be Anthropic’s fault.