Leading artificial intelligence company Anthropic has been directed to immediately halt access to two of its most advanced AI models, Claude Fable 5 and Claude Mythos 5, following an order from the US government. The directive, issued on Friday, cited national security concerns, compelling Anthropic to disable both models for all users globally, not solely foreign nationals as the export control order initially suggested. Anthropic confirmed its compliance but expressed significant frustration and disagreement with the decision.
Mythos 5 stands as Anthropic's most capable AI model, which the company had previewed in early April but kept under tight restrictions due to its exceptional ability to identify security vulnerabilities in software. Anthropic had previously stated that Mythos could find flaws in every major operating system and web browser it tested. Rather than a broad release, the company initiated 'Project Glasswing', sharing the model with approximately 50 vetted organisations, including major tech firms like Amazon, Apple, Google, Microsoft, and cybersecurity specialist CrowdStrike, for defensive cybersecurity applications.
Fable 5, launched only three days prior to the government's order, was Anthropic's commercial response to Mythos. It was designed with enhanced safeguards to block responses in high-risk areas such as cybersecurity and biology, aiming to make it suitable for general public release. Benchmark tests from Vals AI, an AI performance tracking company, had indicated Fable 5 was immediately the most capable AI model available to the public upon its release.
The government's action is framed as an export control measure, intended to restrict foreign national access. However, Anthropic's understanding, detailed in a lengthy blog post, is that the core concern revolves around a claimed 'jailbreak' of Fable 5. The company stated that the government has so far provided only verbal evidence of a 'potential narrow, non-universal jailbreak', which Anthropic describes as prompting the model to read specific code and identify software flaws. Anthropic further contends that this level of capability is already widely accessible in other public models, including OpenAI's GPT-5.5, and is routinely utilised by cybersecurity professionals for defensive purposes.
Anthropic's broader defence highlights that its most robust safeguards operate through independent classifier systems, separate from the model itself. This means that even if a user could bypass a refusal, the underlying protections against dangerous outputs would remain active. The company also noted that a recent review of usage found no evidence of these safeguards being successfully circumvented to produce genuinely harmful content. Anthropic's public statement conveyed strong disagreement, asserting that such a standard, if applied broadly, would effectively halt all new model deployments by frontier AI providers.
The irony of the situation has not been lost on observers. Anthropic has largely built its public identity around being a safety-conscious alternative to its rivals. Its initial caution in restricting Mythos, promoting it as a model too powerful for public release, now appears to have attracted the very government scrutiny that could significantly impact its business operations, particularly as it is widely anticipated to pursue an IPO this year.