Anthropic had to cut off access to two of its AI models late Wednesday, not because of a bug or an outage, but because the US government told it to.
The company received an export control directive from federal officials at 5:21 p.m. Eastern time, ordering it to suspend all foreign national access to Fable 5 and Mythos 5, its two most advanced models. Because the restriction covers foreign nationals both inside and outside the US, including some of Anthropic’s own employees, the company said complying meant disabling both models for every customer rather than trying to sort out access on a case-by-case basis. Anthropic’s other models remain available and unaffected.
The government’s letter reportedly didn’t spell out its specific national security concern in detail, but Anthropic said its understanding is that officials believe someone found a way to bypass, or “jailbreak,” Fable 5’s safety restrictions.
After reviewing a demonstration of the technique, Anthropic said it only surfaced a handful of previously known, fairly minor vulnerabilities, the kind that other publicly available AI models can also expose without needing any special workaround.
Anthropic pushed back hard on the scope of the government’s response. The company said it spent thousands of hours red-teaming Fable’s safeguards before launch, working alongside the US government, the UK’s AI Safety Institute, and several outside research groups, and that no tester found what the industry calls a universal jailbreak, one broad enough to unlock a wide range of dangerous capabilities. What the government reportedly flagged instead was a narrow jailbreak that essentially asks the model to review code and flag security flaws, a capability Anthropic says is already common in tools like OpenAI’s GPT-5.5 and used daily by cybersecurity teams doing legitimate defensive work.
“We disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people,” the company said, adding that applying this standard broadly would likely stall new AI model releases across the entire industry.
Anthropic said it is complying with the order while working to get the suspension lifted, framing the episode as a misunderstanding it hopes to resolve quickly.
