Anthropic Obeys a US Shutdown Order While Publicly Challenging the Basis for It

Tech analyst Varun Mayya begins by explaining that the US government compelled Anthropic to "turn off its strongest models, Mythos and Fable 5, for everyone." The trigger was a national-security letter barring foreign nationals from using those two models anywhere, inside or outside the United States. That restriction also applied to foreign nationals working inside Anthropic.

Running a shared model stack while blocking only specific users was not technically practical. To remain compliant, Anthropic chose to switch both models off entirely, which in turn cut off access for its entire user base, reportedly affecting hundreds of millions of customers.

The government's stated concern was the discovery of a jailbreak method that could bypass Fable 5's safety protections. But when Anthropic examined the technique, it found that it amounted to asking the model to read a specific codebase and fix software defects, something that is already routine with other publicly available models.

The explicit comparison was GPT 5.5. Anthropic argues that the same kind of request is commonly handled by OpenAI's model as well. Its position is that if this level of jailbreak is enough to justify shutting down a model used by hundreds of millions of people, then in practice nearly every major AI release would be vulnerable to the same treatment.

Anthropic took the unusual step of obeying the order while also disputing it in public. Mayya argues that this exposes a category of governance question the industry has not had to answer clearly before: who decides that an AI model is too dangerous to remain online, and what qualifies as sufficient evidence?

Anthropic has also said it believes the decision rests on a misunderstanding and is working to restore access. The video closes by promising follow-up coverage, but the incident may already stand as the first major case in which US national-security authorities exercised immediate shutdown power over a leading AI model provider.