Anthropic's dangerous AI models halted by US export controls

Last week, Anthropic took its new Claude Fable 5 and Mythos 5 AI models offline following a US government export-control directive prohibiting foreign nationals from using them. The company has been in talks with the White House but hasn't secured an agreement to reinstate the offerings. Since Mythos debuted in April, Anthropic has warned of its advanced dual-use capabilities: it can find software vulnerabilities for defenders but also exploit them for attacks. The models were released with limited access—Mythos Preview to a select consortium and Fable 5 publicly with guardrails against biology and cybersecurity queries. The Trump administration believes Fable 5's guardrails can be disabled, posing a national security risk.

Cybersecurity experts argue this clash masks a harder truth: similar capabilities will soon emerge from other companies and open-source models, if they haven't already. Tarah Wheeler of TPO Group noted it's myopic to think no competitor will match Mythos. Bruce Schneier emphasized that smaller, cheaper open-source models with sophisticated prompting can already match performance. An open letter from cybersecurity leaders to the White House called the directive misguided. The real policy challenge, they say, is developing transparent plans for handling advancing AI capabilities in sensitive areas like cybersecurity, rather than focusing on individual models.