The U.S. government ordered Anthropic on Friday to immediately disable access to two of its most advanced AI systems—Claude Mythos 5 and the recently released Claude Fable 5—citing national security concerns, and the company has complied despite publicly disagreeing with the decision. According to Anthropic, the directive arrived at 5:21 p.m. ET and forces the company to cut off both models for every user globally, not just the foreign nationals that export control rules typically target. Other Anthropic models remain unaffected.
The shutdown hits hardest at Mythos 5, which Anthropic considers its most capable model and has kept under tight wraps since previewing it in early April. The company restricted the system because of its unusual ability to identify security vulnerabilities in software—Anthropic says Mythos found flaws in every major operating system and web browser it was tested against. Instead of a broad release, Anthropic launched Project Glasswing, a controlled program giving roughly 50 vetted organizations, including Amazon, Apple, Google, Microsoft, and CrowdStrike, access for defensive cybersecurity work. Fable 5, released just three days before the shutdown, was Anthropic's attempt to bring some of that capability to the public while installing guardrails on high-risk topics like cybersecurity and biology. Benchmark results from Vals AI ranked it as the most capable AI model available to consumers at launch.
The government's stated rationale is export control, but Anthropic says the real concern appears to be a claimed jailbreak of Fable 5. So far, the company reports receiving only verbal evidence of a "potential narrow, non-universal jailbreak"—essentially a prompt that asks the model to scan a specific codebase and flag software weaknesses. Anthropic argues this level of capability is already common in other publicly available models, suggesting the restriction may be both overbroad in scope and weak in evidentiary support. The company has pledged to keep users updated as it works through the implications, but with its flagship public model now offline and its most powerful system locked behind government order, the standoff raises fresh questions about how Washington will draw the line between AI risk and overreach.