US government shut down Anthropic's top AI models over China access fears

Anthropic's two most powerful AI models, Fable 5 and Mythos 5, were abruptly disabled last week after the US government issued an export control directive blocking foreign nationals from accessing them — and the models have stayed offline ever since. The Commerce Department sent a formal letter at 5:21 PM ET on June 12; Anthropic pulled the plug globally within 90 minutes. For enterprise customers — banks, security firms, and government agencies on both sides of the Atlantic — the shutdown arrived without warning or any appeal process.

How it started

Amazon researchers discovered a technique to bypass Fable 5's safety guardrails using specific prompts, allowing users to extract information about cyberattacks. per WSJ, Amazon CEO Andy Jassy personally reported the findings to Treasury Secretary Scott Bessent, which directly triggered the government crackdown. The move puts Amazon in an awkward position: it is both a major investor in Anthropic and the company whose disclosure led to its partner's models being switched off.

The China angle

The official line from White House AI adviser David Sacks focused on the jailbreak itself, without naming any foreign actor. But Semafor reports that the real driver was suspicion that a China-linked group had already accessed Mythos 5 — and that Chinese engineers could potentially reverse-engineer the models for geopolitical purposes. Anthropic disputes this framing, calling the vulnerabilities "minor" and "narrow," and says the severity is comparable to issues already present in publicly available models.

Why the shutdown hit everyone

The White House directive told Anthropic to cut off foreign nationals. That turned out to be technically unworkable on short notice, so Anthropic disabled the models for all users worldwide. UK and European enterprise customers — including financial institutions and security researchers who relied on Mythos for legitimate vulnerability work — lost access alongside everyone else. The UK's AI Safety Institute had participated in red-teaming Fable 5 before launch but was not consulted ahead of the shutdown, exposing a real gap in US-UK coordination on AI governance.

Anthropic says the models remain offline while it works toward a compliance solution. No timeline for reinstatement has been given, and no appeals path has been made public.