U.S. Government Orders Anthropic to Disable Claude Mythos 5 and Fable 5 Over Jailbreak Concerns

The U.S. government on Friday, June 13, ordered Anthropic to immediately shut down access to Claude Mythos 5 and Claude Fable 5 worldwide, citing national security concerns. According to TechCrunch AI, Anthropic received the directive at 5:21 PM ET and announced compliance via X, but the company made clear in a lengthy blog post that it believes the government’s action is unjustified. The order is framed as an export control action nominally restricting foreign national access, but the underlying trigger is a claimed vulnerability in Fable 5 that Anthropic disputes both the severity and novelty of.

Mythos 5’s Restricted Prowess: Project Glasswing and the Security Dilemma

Claude Mythos 5, which Anthropic previewed in early April, represents the company’s most capable model—one so powerful that Anthropic kept it tightly restricted out of concern over its exceptional ability to identify security vulnerabilities in software. According to TechCrunch AI, Mythos identified flaws in every major operating system and web browser it tested during internal evaluation. Rather than release it broadly, Anthropic launched Project Glasswing, a controlled program sharing the model with approximately 50 vetted organizations including Amazon, Apple, Google, Microsoft, and CrowdStrike for defensive cybersecurity work only.

This restraint reflected Anthropic’s public safety posture—the company had warned repeatedly about the risks of deploying such a capable vulnerability-finding tool publicly. Yet this very caution, paradoxically, may have hastened the government intervention by highlighting capabilities that regulators viewed as dangerous.

Fable 5’s Failed Gambit: Guardrails and the Jailbreak Question

Three days before the shutdown order, Anthropic released Fable 5 as a commercial answer to market pressure. According to TechCrunch AI, Fable 5 was Mythos 5 retrofitted with guardrails designed to block high-risk outputs in cybersecurity and biology, making it safe enough for general release. Performance data from Vals AI indicated it was immediately the most capable publicly available model. But the government’s concern centers on what it characterizes as a jailbreak of those guardrails.

Anthropic disputes the framing. The company states the government has provided only verbal evidence of a “potential narrow, non-universal jailbreak”—one amounting to prompting the model to read a specific codebase and identify software flaws. Anthropic argues this capability is already widely available in competing public models, including OpenAI’s GPT-5.5, and is routinely used by cybersecurity professionals for defensive purposes. The company also contends its safeguards operate through independent classifier systems functioning separately from the model, meaning even a jailbroken conversation would not bypass the underlying protections against the most dangerous outputs.

Why This Matters

The shutdown creates immediate friction for enterprise AI procurement teams: they now face a narrowed choice between restricted-access models (like Mythos 5 through Project Glasswing) and public models subject to sudden regulatory recall. Organizations that based Q3 vendor evaluations on Fable 5’s benchmark performance will need to reassess. More broadly, the order signals that the U.S. government is willing to mandate retroactive model shutdowns based on narrowly reported vulnerabilities, not just export restrictions—a precedent that could reshape how AI companies weigh safety disclosure against regulatory preemption. If competitors and other regulators adopt similar thresholds, the economics of releasing frontier models may shift fundamentally toward longer restricted-access phases, delaying public availability and competition.