White House export controls shut down Anthropic's Fable 5 over disputed jailbreak
The Trump administration issued export controls against Anthropic's Fable 5 and Mythos 5 models over a disputed jailbreak, forcing Anthropic to shut them down for all customers — including American ones.
Score breakdown
The export controls effectively removed what the article describes as perhaps the most powerful AI in the world from use by American companies and the U.S. government itself, while leaving comparable models from other providers unaffected — a concrete outcome that cuts against the administration's stated goal of winning the AI race against China.
- 01The White House gave Anthropic only 90 minutes to voluntarily take down Fable 5 and Mythos 5 before issuing export controls.
- 02Fable 5 is a public release of Claude Mythos Preview, which Anthropic had originally restricted to a small group of cybersecurity partners.
- 03The export controls bar any foreign national from using the models, including Anthropic employees inside the United States.
The Trump administration, which had publicly committed to accelerating U.S. AI development to beat China, moved against Anthropic's newest and most capable models in a move the article by Matteo Wong in The Atlantic frames as self-defeating. Anthropic had just released Fable 5 — a public version of Claude Mythos Preview, a model the company had previously restricted to a small group of cybersecurity partners due to its hacking capabilities. Before the public release, Anthropic developed safety guardrails that underwent third-party testing, including with the U.S. government. After Fable 5 launched, cybersecurity experts largely criticized the model for being too restrictive, not too dangerous.
The administration's reversal came after Amazon researchers alerted officials to a potential way to circumvent Fable 5's safety systems.
The administration's reversal came after Amazon researchers alerted officials to a potential way to circumvent Fable 5's safety systems. Officials gave Anthropic a 90-minute window to voluntarily take down Fable 5 and Mythos 5 — a newer, limited-release version of Mythos Preview. When Anthropic declined, the government issued export controls, a designation that prevents foreign nationals from using the models even if employed by Anthropic within the United States. To comply rapidly, Anthropic shut the models down for all customers, including American ones and the U.S. government itself.
The severity of the underlying jailbreak is sharply contested. A White House spokesperson described it as "very serious" but said details are classified. Anthropic characterized the elicited actions as "either entirely benign responses or are minor findings." Cybersecurity expert Katie Moussouris, CEO of Luta Security, reviewed the White House's report at Anthropic's request and described the behavior as Fable refusing one prompt but complying with a rephrased version — standard cyberdefense functionality. She noted that OpenAI's GPT-5.5, which she said has similar cybersecurity capabilities, faces no such controls, nor do less advanced Anthropic models like Opus 4.8 that can perform many of the same tasks. Alex Stamos, chief security officer at AI-coding company Corridor, told the article's author that the jailbreak did not appear to elicit the advanced cyber abilities that made Mythos notable.
Key facts
- 01The White House gave Anthropic only 90 minutes to voluntarily take down Fable 5 and Mythos 5 before issuing export controls.
- 02Fable 5 is a public release of Claude Mythos Preview, which Anthropic had originally restricted to a small group of cybersecurity partners.
- 03The export controls bar any foreign national from using the models, including Anthropic employees inside the United States.
- 04Anthropic shut down Fable 5 and Mythos 5 for all customers — including American ones — to rapidly comply.
- 05The jailbreak was flagged to the administration by researchers at Amazon.
- 06Cybersecurity expert Katie Moussouris reviewed the White House's report and described the behavior as 'the model working as intended' for cyberdefense.
- 07OpenAI's GPT-5.5 and Anthropic's Opus 4.8, described as having similar capabilities, are not subject to the same export controls.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 16, 2026 · 23:11 UTC. How this works →