Anthropic Pulls AI Models After US Export-Control
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Anthropic took Claude Fable 5 and Mythos 5 offline last week after a US export-control directive barred "any foreign national" from using them, and has been negotiating with the White House since Friday without reaching a deal to reinstate the models.
- The Trump administration restricted the models on the stated rationale that Fable 5's guardrails can be disabled to access Mythos 5's full capabilities, allegedly creating a national security risk.
- Anthropic warned since Mythos debuted in April that the model is dual-use — capable of both finding software vulnerabilities for defenders and developing exploits — initially releasing Mythos Preview to a consortium called Project Glasswing while releasing Fable 5 to the public with blocks on biology and cybersecurity queries.
- OpenAI also privately released a cybersecurity-focused model in mid-April and announced an expanded cybersecurity strategy, illustrating that the capability is not unique to Anthropic.
- Tarah Wheeler of TPO Group called it "myopic in the extreme" to think competitors lack similar capabilities, saying other companies are "hot on Anthropic's heels" and holding models in reserve to see how regulators treat Anthropic first.
- Anthropic's Logan Graham told WIRED in April that the real message is preparing for a world where these capabilities are broadly available within 6 to 24 months — a timeline echoed by Bruce Schneier, who argued smaller open-source models can already match Mythos/Fable with sophisticated prompting.
- A large group of cybersecurity leaders sent the administration an open letter on Sunday calling the export-control directive misguided, and Veracode cofounder Chris Wysopal framed the core question as whether such restrictions meaningfully reduce risk or merely slow down the people trying to make systems safer.
Why it matters: The administration cast the move as preventing foreign access to dangerous AI, but a coalition of cybersecurity researchers argues that targeting one company's models is futile because competitors and open-source developers will reach equivalent capabilities within months — potentially handicapping the very defenders the policy claims to protect. The open letter from cybersecurity leaders on Sunday calling the directive "misguided" is the substantive institutional pushback, not corporate complaining.
