Anthropic’s Safety Warnings Backfire as Government Halts Its Most Powerful AI

On Friday, the U.S. government mandated that Anthropic immediately disable access to two of its most advanced AI systems, Claude Fable 5 and Claude Mythos 5, citing national security risks. Anthropic confirmed compliance via X but stated it believes the government’s decision was misguided.

According to Anthropic, the directive arrived at 5:21 pm ET on Friday, requiring the company to block both models for all global users, extending beyond the foreign nationals originally targeted by the export control order. Other Anthropic models remain unaffected.

Why does this matter? Mythos represents Anthropic’s most powerful AI, previewed in early April and kept under strict restriction due to its exceptional capacity to identify software security vulnerabilities. Anthropic reports that Mythos found flaws in every major operating system and web browser tested. Consequently, instead of a broad release, the company launched Project Glasswing, a controlled program distributing access to approximately 50 vetted organizations, including Amazon, Apple, Google, Microsoft, and CrowdStrike, for defensive cybersecurity purposes.

Fable 5, launched just three days ago, addresses commercial pressure by offering a version of Mythos with guardrails that block responses in high-risk sectors like cybersecurity and biology. Anthropic argues this makes it safe for general use, and benchmark tests by Vals AI, which tracks AI performance, confirm it is currently the most capable public AI model.

Image Credits:Vals AI /

While the government frames this as an export control measure limiting foreign access, Anthropic argues in a detailed blog post that the real issue is a claimed jailbreak of Fable 5. The company states the government has only provided verbal evidence of a “potential narrow, non-universal jailbreak,” which involves prompting the model to analyze a specific codebase for software flaws. Anthropic notes this capability is already widely available in other public models, such as OpenAI’s GPT-5.5, and is routinely used by cybersecurity professionals for defense.

Anthropic’s core argument is that its strongest safeguards rely on independent classifier systems separate from the model itself. This means that even if a user bypasses a refusal, underlying protections against dangerous outputs remain active.

Despite these safeguards, the government proceeded with its action, and Anthropic has openly expressed its frustration. “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,” the company wrote. “If this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers.”

Anthropic is widely expected to pursue an IPO this year, positioning itself as the safety-conscious alternative to competitors. Observers note the irony that the caution Anthropic exercised in restricting Mythos — promoting it as too dangerous for public release — has now triggered the exact government scrutiny that could disrupt its business.

OpenAI’s Sam Altman appears to be enjoying the situation. In April, he told podcaster Ashlee Vance that Anthropic’s handling of Mythos was “fear-based marketing.” “It is clearly incredible marketing to say, ‘We have built a bomb. We were about to drop it on your head. We will sell you a bomb shelter for $100 million,’” Altman said. While Altman, whose company also plans an imminent IPO, did not predict a government shutdown, he highlighted a consequence of Anthropic’s strategy: by spending months warning that its AI is uniquely dangerous, the company ensured that the U.S. government would take notice.