Overview
- Anthropic said it received an export‑control directive from the U.S. government on Friday that required blocking access to Fable 5 and Mythos 5 for any foreign national, and the company disabled both models for all users to ensure compliance.
- The government’s concern, as described by Anthropic, is a demonstrated technique that can bypass Fable 5’s safeguards by asking the model to read a specific codebase and fix software flaws; Anthropic says officials provided only verbal evidence and called the finding a narrow, non‑universal jailbreak.
- Anthropic built Fable 5 with classifiers that flag queries about cybersecurity, biology, chemistry and model development and automatically route risky requests to a less capable model, Claude Opus 4.8, but the company says no universal jailbreak has been found.
- Security researchers and Anthropic warn Mythos‑class models can speed the conversion of disclosed bugs into working exploits in hours, raising dual‑use concerns that the government cited in its national‑security rationale.
- The directive revives wider policy and geopolitical questions about treating frontier AI as a strategic asset, complicates Anthropic’s tense relations with U.S. agencies after a Pentagon supply‑chain designation, and could reshape who can use top AI models worldwide.