On June 12, Anthropic received a directive from the White House: immediately block foreign nationals from accessing its newly released models, Fable 5 and Mythos 5. The order, tied to national security concerns over potential cyberattack capabilities, forced the company to take both models offline for all customers—including its own foreign-born researchers.

The backstory traces to conversations between Amazon CEO Andy Jassy and the administration. Amazon researchers had demonstrated a series of prompts that could coax Fable 5 into returning information useful for cyberattacks. The White House acted swiftly, citing risks of foreign access and potential model distillation—where a smaller model is trained to mimic a larger one, effectively stealing its capabilities.

Anthropic complied but pushed back publicly. In a statement, the company argued the government relied on verbal descriptions of a narrow jailbreak, not a documented vulnerability, and noted that similar weaknesses exist in other frontier models like GPT-5.5. The company described the exploit as minor and said it disagreed that a single jailbreak path warranted recalling a model deployed to hundreds of millions of users.

The incident exposes a tension at the heart of US AI dominance: the same government that enables frontier research can also unilaterally restrict access, even for domestic teams. For engineers, the episode raises practical questions about deployment architecture—how do you design model serving infrastructure that can rapidly enforce region-level access controls without breaking your own CI/CD pipeline? And what happens to your training data pipelines when your own staff can’t query the model?

Anthropic’s decision to cut all access, rather than implement granular nationality checks, suggests the technical cost of compliance was lower than the risk of non-compliance. For now, Fable 5 and Mythos 5 remain dark, and the industry watches for precedent.

Source: The Verge