The US government just unplugged Claude's best models
On Friday June 13, the Trump administration ordered Anthropic to halt all access to Fable and Mythos 5 for foreign nationals everywhere, including outside the United States and including Anthropic's own employees, citing national security. Anthropic disabled the models. Six days after we covered Fable 5's public launch, the government decided the public model was too dangerous to let out and pulled the plug.
The trigger was a claimed jailbreak. Anthropic said the export-control order may have been based on a report of a method to bypass Fable's safeguards and unlock its Mythos-level capabilities, the bio and cyber reasoning the safety layer is supposed to fence off. According to Katie Moussouris, who signed the protest letter, the method came from an Amazon research paper that didn't actually demonstrate a real jailbreak. Reporting has tied it to three words, fix this code, as the reasoning thread that allegedly walks the model toward the dangerous capability.
Then the cybersecurity community pushed back hard. By June 15 an open letter signed by about 100 professionals, from Nvidia, Adobe, Zoom, Google, Anaplan, and Sophos plus academic researchers, asked the government to lift the order. Their argument is the inversion that makes export controls on defensive AI so strange: this action has taken the best models away from defenders who use them to find vulnerabilities and harden software, while attackers, who don't follow export controls, lose nothing.
For anyone building on frontier models, this is the concrete new risk. If you're a non-US national, or you employ any, your access to the top Anthropic models can vanish by executive order overnight. Model availability is now a sovereignty question, not just a pricing one. And it knots straight into the safety thread we've been pulling, where Fable 5's careful bio-cyber fallback meets Hades-style attacks on the safety layer itself. The frontier just hit the policy wall, in public, with a hundred defenders saying the wall is pointed the wrong way.
← Back to all articles
The trigger was a claimed jailbreak. Anthropic said the export-control order may have been based on a report of a method to bypass Fable's safeguards and unlock its Mythos-level capabilities, the bio and cyber reasoning the safety layer is supposed to fence off. According to Katie Moussouris, who signed the protest letter, the method came from an Amazon research paper that didn't actually demonstrate a real jailbreak. Reporting has tied it to three words, fix this code, as the reasoning thread that allegedly walks the model toward the dangerous capability.
Then the cybersecurity community pushed back hard. By June 15 an open letter signed by about 100 professionals, from Nvidia, Adobe, Zoom, Google, Anaplan, and Sophos plus academic researchers, asked the government to lift the order. Their argument is the inversion that makes export controls on defensive AI so strange: this action has taken the best models away from defenders who use them to find vulnerabilities and harden software, while attackers, who don't follow export controls, lose nothing.
For anyone building on frontier models, this is the concrete new risk. If you're a non-US national, or you employ any, your access to the top Anthropic models can vanish by executive order overnight. Model availability is now a sovereignty question, not just a pricing one. And it knots straight into the safety thread we've been pulling, where Fable 5's careful bio-cyber fallback meets Hades-style attacks on the safety layer itself. The frontier just hit the policy wall, in public, with a hundred defenders saying the wall is pointed the wrong way.
Comments