Anthropic's Safety Warnings Backfire: Government Halts Its Most Powerful AI

On Friday, the U.S. government issued a directive to Anthropic, mandating the immediate cessation of access to two of its most potent AI models, Claude Fable 5 and Claude Mythos 5, citing national security implications. Anthropic confirmed its compliance via an announcement on X, though it simultaneously expressed disagreement with the government's decision.

Anthropic reported receiving this directive at 5:21 pm ET on Friday. It compels the company to deactivate both models for all global users, extending beyond the foreign nationals that the government's export control order ostensibly targeted. The accessibility of Anthropic's other AI models remains undisturbed.

The significance of this decision stems from Mythos being Anthropic's most advanced AI model. Previewed in early April, it has been kept under stringent restrictions due to its remarkable proficiency in uncovering software security vulnerabilities, as described by Anthropic. The company stated that Mythos successfully identified flaws across every major operating system and web browser it evaluated. Consequently, instead of a broad release, Anthropic initiated Project Glasswing, a controlled program that granted access to approximately 50 vetted entities, including industry giants like Amazon, Apple, Google, Microsoft, and CrowdStrike, for use in defensive cybersecurity applications.

Fable 5, introduced merely three days prior, represented Anthropic's response to clear commercial demands. It was presented as a version of Mythos equipped with protective guardrails designed to prevent outputs in high-risk domains such as cybersecurity and biology, thereby rendering it suitable for general public release, the company asserted. Benchmark assessments conducted by Vals AI, an organization specializing in AI technology performance tracking, immediately recognized Fable 5 as the most capable AI model accessible to the public.

While the government's directive is characterized as an export control measure, limiting access for foreign nationals, Anthropic's detailed blog post indicates its perception that the root concern is an alleged "jailbreak" of Fable 5. According to Anthropic, the government has, to date, offered only verbal evidence of a "potential narrow, non-universal jailbreak." The company clarifies this as a process where the model is prompted to analyze a specific codebase and pinpoint software vulnerabilities. Anthropic further contends that this "level of capability" is already broadly accessible in other publicly available models, including OpenAI's GPT-5.5, and is routinely employed by cybersecurity experts for defensive operations.

Anthropic's overarching argument emphasizes that its most robust safeguards are implemented via independent classifier systems, operating distinctly from the core model. This architecture implies that even if a user were to prompt Fable to override a refusal, the fundamental protections against highly dangerous outputs would persist. Furthermore, the company highlighted in its post that a recent review of usage patterns revealed no instances of these safeguards being successfully circumvented to generate genuinely harmful content.

Evidentially, these arguments were insufficient to deter government intervention, and Anthropic has openly expressed its frustration. 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." It further cautioned, "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 broadly anticipated to pursue an initial public offering (IPO) this year, having largely cultivated its public image as a safety-focused competitor. Observers note the irony that the very prudence Anthropic demonstrated in restricting Mythos—a model it characterized as too dangerous for public release—has seemingly now drawn the precise governmental oversight that could most significantly impede its commercial operations.

Sam Altman of OpenAI is likely finding some satisfaction in these developments. In April, he remarked to podcaster Ashlee Vance that Anthropic's approach to Mythos constituted "fear-based marketing." Altman elaborated, stating, "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.'" While Altman, whose company is also widely expected to launch an IPO imminently, did not foresee a government shutdown, he did pinpoint a factor that has now seemingly impacted Anthropic: when a company spends months asserting that its AI is uniquely perilous, the world — including the U.S. government — tends to heed such warnings.