On June 11, 2026, at 5:21pm ET, Anthropic received a letter from the US government ordering them to immediately shut down access to their two most powerful AI models. No appeal window. No negotiation. Just comply now.
They did. And then they published a statement saying the government got it wrong.
What the government actually said
The directive came from national security authorities and cited export control law. The core claim: someone found a way to jailbreak Fable 5, Anthropic’s flagship model, and the government believes that poses a risk to national security.
Because the restriction targeted any “foreign national,” and because Anthropic has foreign national employees and customers all over the world, the only practical way to comply was to kill access for everyone. Hundreds of millions of users lost access overnight.
The letter did not explain what specific threat was identified. No written technical evidence was provided. Anthropic says the government gave them only verbal evidence of the potential jailbreak.
Anthropic’s case against the decision
Anthropic is not staying quiet about this. Their public statement is direct and worth reading in full, but here is the core of their argument.
The jailbreak the government flagged is narrow and non-universal. What does that mean? A non-universal jailbreak can get a model to do something it normally wouldn’t in a specific scenario, but it cannot broadly bypass the model’s safety systems across a wide range of harmful tasks. Anthropic said clearly at launch that this type of jailbreak is likely unavoidable for any model in the industry today.
They also checked whether the same jailbreak works on competing models. It does. According to their statement, the technique produces results already available from OpenAI’s GPT-5.5 — a model that is still fully accessible to users worldwide.
“We have not even received a disclosure of a concerning non-universal potential jailbreak that led to a harmful result.” — Anthropic, June 11, 2026
The specific technique flagged? Asking the model to read a codebase and fix security flaws. That is a task security professionals use every day to defend systems, not attack them.
Why this matters beyond Anthropic
Anthropic did not just push back on this specific order. They pushed back on the process.
Their statement argues that if a narrow, non-universal jailbreak is enough to trigger a recall of a commercial model, then no frontier AI company could ever ship a new model. Every major model deployed today has some version of this vulnerability. That includes models from Google, Meta, and OpenAI.
The government’s directive also arrived with no written technical justification, no transparent statutory process, and no right of appeal before enforcement. That is not just bad for Anthropic. It sets a standard where any model, from any company, can be pulled at any time based on verbal claims that cannot be challenged.
Anthropic has said it supports government oversight of unsafe AI deployments. But they want that oversight to be transparent, fair, and grounded in technical facts. This action, they say, was none of those things.
What happens next
Anthropic says it is complying fully with the directive while working to restore access. They plan to share more technical details about the jailbreak within 24 hours of the June 11 announcement.
All other Anthropic models, including Claude, remain fully available. Only Fable 5 and Mythos 5 are affected.
If you rely on Fable 5 or Mythos 5 in your workflow, check Anthropic’s official statement page for the latest updates on when access might be restored.
The bigger question here is not about one model getting pulled. It is about whether the US government will build a real, technical, transparent framework for AI oversight before it needs to make the next call like this one. Right now, the answer looks like no.
