By Harry Sekulich – June 13, 2026
Anthropic has paused the rollout of its newest AI system, Claude Fable 5, after a U.S. government order warned that the model could be jailbroken to bypass safety filters.
The company’s statement on its website says the order “requires us to cancel Fable 5 and Mythos 5 for all customers to ensure compliance.” The directive comes days after the public launch of the model, which has been marketed as a rival to OpenAI’s ChatGPT and Google’s Gemini.
Anthropic noted that no specific security flaws had been identified by authorities, but the U.S. intelligence community believes they can find ways to bypass or “jailbreak” the system. Jailbreaking is a process that circumvents software barriers to access sensitive data or unlock hidden features.
In a review of a demonstration, the company said it found a handful of minor vulnerabilities that could, in theory, be exploited by the jailbreaking technique. However, it argues these weaknesses were simple and could be discovered by other publicly‑available models without the need for a bypass.
Prior to release, Anthropic granted a limited pre‑launch of Claude Fable 5 to a handful of organisations for testing and to catch any issues. The company explained that the AI’s intelligence was “too powerful to release” until it was fully vetted.
The incident has intensified scrutiny from finance and technology leaders. Trump’s administration took the case public, with Defense Secretary Pete Hegseth labeling Anthropic a “supply‑chain risk,” a designation usually reserved for companies based in adversarial countries.
Anthropic is suing the Pentagon over the prohibition that bars U.S. government agencies from using its tools. A federal judge has ruled that the directive cannot be enforced when the lawsuit is pending, allowing agencies to continue using Anthropic’s services.
The pause signals a broader debate on AI governance, revealing how quickly a new tool can spark legal and regulatory backlash when security concerns are raised, and prompting a reevaluation of how AI systems are tested and released.
