The U.S. government on Friday ordered Anthropic to immediately cut off access to two of its most powerful AI models, Claude Fable 5 and Claude Mythos 5, citing national security concerns. According to TechCrunch AI, the company complied right away but made clear it thinks the government got this one wrong. The directive landed at 5:21 pm ET and forced Anthropic to disable both models for every user worldwide, not just the foreign nationals the export control order nominally targeted. Access to Anthropic’s other models wasn’t touched.
What stands out here is the irony. Anthropic spent months telling the world these models were uniquely dangerous, and now that warning may be the very thing that brought down the hammer.
What got shut off
Two models, two very different stories:
- Mythos 5 is Anthropic’s most capable model, previewed in early April and kept tightly restricted ever since. The company says it can find security vulnerabilities in software at an exceptional level, flagging flaws in every major operating system and web browser it tested. Rather than release it broadly, Anthropic ran a controlled program called Project Glasswing, sharing it with roughly 50 vetted organizations including Amazon, Apple, Google, Microsoft, and CrowdStrike for defensive cybersecurity work.
- Fable 5 shipped just three days ago. It’s Mythos fitted with guardrails that block responses in high-risk areas like cybersecurity and biology, built for general release. Per benchmark tests from Vals AI cited by TechCrunch AI, it was immediately the most capable AI model available to the public.
The government’s case, and Anthropic’s pushback
The directive is framed as an export control action restricting foreign access. But Anthropic says its read is that the real trigger is a claimed jailbreak of Fable 5. So far, the company says, the government has offered only verbal evidence of a “potential narrow, non-universal jailbreak,” which amounts to prompting the model to read a specific codebase and identify software flaws.
Anthropic’s counterargument is blunt. That level of capability is already widely available in other public models, including OpenAI’s GPT-5.5, and cybersecurity pros use it routinely for defensive work. The company also notes its strongest safeguards run through independent classifier systems that sit separately from the model, so even if someone talks Fable past a refusal, the protections against the most dangerous outputs stay in place.
“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. It warned that if that standard were applied across the industry, “it would essentially halt all new model deployments for all frontier model providers.”
Why this matters
This is significant for three reasons:
- It sets a precedent. A verbal report of a narrow jailbreak was enough to pull a flagship commercial model worldwide. Every frontier lab is watching how that line gets drawn.
- The timing stings. Anthropic is widely expected to pursue an IPO this year and has built its identity as the safety-conscious alternative to rivals. The caution it marketed as a virtue is now the thing disrupting its business.
- Marketing has consequences. When you spend months calling your own model uniquely dangerous, regulators tend to take you at your word.
OpenAI’s Sam Altman saw this coming, sort of. Back in April he called Anthropic’s handling of Mythos “fear-based marketing,” comparing it to saying “we have built a bomb” and then selling you the bomb shelter. He didn’t predict a government shutdown, but he flagged the dynamic that just bit Anthropic.
What to watch next
The immediate question is how long the shutoff lasts and whether the government produces written, specific evidence to back the order. If Anthropic can’t get a fast reversal, expect the dispute to spill into public and possibly legal channels, especially with an IPO on the line. The bigger story is the new reality every lab now faces: the gap between previewing a powerful model and keeping it switched on may come down to a single regulator’s phone call. More details are available at the original TechCrunch AI report.