Overview
- Anthropic said it received an export‑control directive from the U.S. government and on Saturday disabled Fable 5 and Mythos 5 for all customers because it could not practically block access by non‑U.S. nationals.
- The government flagged a reported method to 'jailbreak' Fable 5 and bypass its safety guards so the model could be used to find software vulnerabilities, but authorities have not publicly released detailed technical evidence.
- Anthropic disputes the scope of the order, says it reviewed the cited report and found only a limited technique that exposed a small number of known, minor bugs that other public models can also detect, and called the measure a 'misunderstanding.'
- As it seeks to restore access, Anthropic is changing safeguards by tuning cyber and bio classifiers, making fallbacks visible (routing risky prompts to Opus 4.8) and returning API rejection reasons, while warning these steps may raise short‑term false positives.
- The episode creates a new precedent for treating access to a commercial AI model as an export‑controlled item and sharpens broader disputes over transparency, independent audits, and how regulators should limit high‑capability models.