Anthropic’s safety warnings may have just backfired — the government has pulled the plug on its most powerful AI

In a surprising move, the U.S. government mandated Anthropic to immediately suspend access to its two most advanced AI systems, Claude Fable 5 and Claude Mythos 5, citing national security risks. Anthropic announced compliance with the directive on social media, expressing its belief that this decision is misguided. The order was communicated to Anthropic at 5:21 PM ET on Friday, compelling the company to deactivate these models for all global users. This directive extends beyond the foreign nationals targeted by the government’s export control measures. Fortunately, access to Anthropic's other AI models remains unaffected. The implications of this action are significant. Mythos, Anthropic's most sophisticated AI, was introduced earlier this year and has been closely monitored due to its unique capability to uncover security vulnerabilities in software. Anthropic claims that Mythos successfully identified flaws in every major operating system and web browser it evaluated. Rather than releasing it to the public, the company opted for a controlled release through Project Glasswing, collaborating with about 50 vetted organizations, including tech giants like Amazon, Apple, and Google, for cybersecurity purposes. On the other hand, Fable 5, launched just three days prior, was designed to address commercial pressures, featuring safeguards to limit responses in sensitive areas like cybersecurity and biology. This model was touted as the most advanced publicly available AI, based on evaluations from Vals AI, a company that tracks AI technology performance. The government’s directive is officially framed as an export control measure to restrict access for foreign nationals. However, Anthropic suggests that the underlying issue stems from a reported jailbreak of Fable 5. The company states that the government has thus far provided only verbal evidence of a “potential narrow, non-universal jailbreak,” which it describes as a scenario where the model is prompted to analyze a specific codebase for software vulnerabilities. Anthropic points out that similar capabilities are already accessible in other widely used models, including OpenAI’s GPT-5.5. The company also emphasizes that its stringent safeguards operate through independent classifiers that are separate from the model itself. This means that even if someone manages to bypass the model's refusal to respond, the core protections against harmful outputs would still be intact. Moreover, a review of recent usage revealed no successful breaches of these safeguards. Despite these arguments, the government proceeded with its action, prompting frustration from Anthropic. The company stated, “We disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people.” They asserted that applying such a standard across the industry would effectively halt the deployment of new models by leading providers. As Anthropic prepares for a potential IPO later this year, the irony is palpable: the very caution it exercised with Mythos, promoting it as too dangerous for public release, has drawn the scrutiny that threatens its operations. Observers note that this situation may provide some satisfaction to OpenAI's Sam Altman, who previously critiqued Anthropic's marketing strategy, suggesting it was based on fear tactics. Altman remarked that portraying their AI as a catastrophic threat could backfire, as it has seemingly done in this instance.