Government Pulls Plug on Anthropic’s Most Powerful AI After Safety Warning Backfires
▼ Summary
– The U.S. government ordered Anthropic to immediately disable its Claude Fable 5 and Claude Mythos 5 models worldwide, citing national security concerns; Anthropic complied but disagreed with the decision.
– Mythos 5, Anthropic’s most capable model, was kept tightly restricted due to its exceptional ability to find security flaws in major operating systems and browsers, shared with only 50 vetted organizations for defensive use.
– Fable 5 was a publicly released version of Mythos with guardrails blocking high-risk areas like cybersecurity and biology, and was immediately the most capable public model based on benchmark tests.
– Anthropic argues the government’s concern is based on a verbal claim of a “potential narrow, non-universal jailbreak” of Fable 5, which the company says represents capability already available in other models like OpenAI’s GPT-5.5.
– Anthropic noted that its safeguards operate through independent classifiers and a review found no evidence of successful bypasses producing harmful content, but the government acted anyway, leading Anthropic to warn that applying this standard would halt all new model deployments.
The U.S. government issued an emergency directive on Friday, ordering Anthropic to immediately disable access to its two most powerful AI models, Claude Fable 5 and Claude Mythos 5, citing national security concerns. Anthropic confirmed on X that it has complied with the order, though the company expressed strong disagreement with the decision, arguing the government’s assessment is flawed.
The directive, received by Anthropic at 5:21 pm ET on Friday, requires the company to block access to both models for all users worldwide, not just the foreign nationals that the export control order was technically designed to target. Other Anthropic models remain unaffected by this action.
The stakes are high because Mythos represents Anthropic’s most advanced AI system, first previewed in early April and kept under tight restrictions ever since. Anthropic has described Mythos as having an extraordinary ability to identify security vulnerabilities in software, successfully finding flaws in every major operating system and web browser it tested. Rather than releasing it widely, the company launched a controlled program called Project Glasswing, sharing access with approximately 50 vetted organizations including Amazon, Apple, Google, Microsoft, and CrowdStrike for defensive cybersecurity purposes.
Fable 5, released just three days before the government’s order, was Anthropic’s commercially viable answer to the pressure surrounding Mythos. The company fitted it with guardrails designed to block responses in high-risk areas like cybersecurity and biology, arguing this made it safe for general release. According to benchmark tests from Vals AI, a firm tracking AI performance, Fable 5 immediately became the most capable AI model available to the public.
Though the government frames its directive as an export control measure restricting foreign national access, Anthropic revealed in a detailed blog post that the underlying concern appears to be a claimed jailbreak of Fable 5. The company says the government has provided only verbal evidence of what it describes as a “potential narrow, non-universal jailbreak.” According to Anthropic, this jailbreak amounts to prompting the model to read a specific codebase and identify software flaws, a capability the company argues is already available in other publicly accessible models, including OpenAI’s GPT-5.5. Anthropic also notes that cybersecurity professionals routinely use such capabilities for defensive purposes.
Anthropic’s defense rests on the fact that its strongest safeguards operate through independent classifier systems separate from the model itself. Even if someone convinces Fable to continue past a refusal point, the company argues, the underlying protections against the most dangerous outputs remain intact. A review of recent usage found no evidence of those safeguards being successfully bypassed to produce genuinely harmful content.
None of this prevented the government from acting, and Anthropic has not hidden 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.”
The timing is particularly awkward for Anthropic, which is widely expected to pursue an IPO this year and has built its public identity around being the safety-conscious alternative to rivals. Observers note the irony that Anthropic’s caution in restricting Mythos, which it promoted as a model so dangerous it couldn’t be publicly released, has now attracted the exact kind of government scrutiny that could most disrupt its business.
OpenAI’s Sam Altman must be watching with some satisfaction. In April, he told podcaster Ashlee Vance that Anthropic’s handling of Mythos amounted to “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 didn’t predict a government shutdown, he identified a dynamic that has now come back to bite Anthropic: when you spend months telling the world your AI is uniquely dangerous, the world, including the U. S. government, tends to listen.
(Source: TechCrunch)
