Anthropic Pulls Two Models After Government Export Order Cites Jailbreak Risk

Anthropic Pulls Two Models After Government Export Order Cites Jailbreak Risk

Anthropic took Claude Fable 5 and Mythos 5 offline Friday evening, complying with a US government export control directive issued at 5:21 PM ET. The order, citing national security concerns, required Anthropic to block access for any foreign national—including its own employees. To ensure compliance, the company cut access for all customers globally.

This is the first time a US export order has targeted a publicly released AI model. The government’s stated concern: a jailbreak method that could bypass Fable 5’s safeguards on cybersecurity, biology, and chemistry topics. Anthropic’s blog post notes the government demonstrated the technique identifying “a small number of previously known, minor vulnerabilities” that other publicly available models can also find without a bypass.

The incident escalates tensions between Anthropic and the Trump administration. Earlier this year, the Department of Defense labeled Anthropic a “supply chain risk” after the company tried to restrict military use of its technology. Anthropic responded with lawsuits.

Fable 5, released Tuesday, was a safety-hardened version of Mythos. The Mythos Preview had a limited April rollout to let organizations test its cybersecurity defense capabilities. Anthropic said the public release was coordinated with the US government.

From a systems perspective, the engineering trade-off is stark: Fable 5’s narrow jailbreak—essentially prompting the model to read and fix a specific codebase—does not appear to grant attackers capabilities beyond what other models provide. Yet the government’s order applies no such nuance. For data and ML teams, this raises questions about deployment architectures that can support rapid, jurisdiction-aware access revocation without full service interruption. Anthropic’s all-or-nothing shutdown suggests their current access control layer lacks that granularity.

CEO Dario Amodei reiterated support for a transparent government review process for unsafe models. Friday’s action, Anthropic argued, “does not adhere to those principles.”

Source: Wired

Source:Wired ↗
← Back to News