Overview
- Anthropic restored global access to Claude Fable 5 after the Commerce Department lifted export controls on June 30 and redeployed the model with a classifier that flags cybersecurity‑related prompts.
- Mythos 5 remains restricted to a vetted set of defensive partners under Project Glasswing while Anthropic and U.S. agencies run joint pre‑release testing and evaluations.
- When the classifier blocks a flagged request it is redirected to the smaller Opus 4.8 model and Anthropic says the classifier stops the reported bypass technique in more than 99% of cases.
- The export‑control intervention followed reports that researchers had found a prompt‑based way to bypass Fable’s guardrails and agency concern about Mythos’s ability to autonomously probe systems, prompting a negotiated staged redeployment and a HackerOne reporting program.
- Policy uncertainty from the episode is already pushing some customers toward open‑weight and non‑U.S. models and is accelerating talks on shared industry standards for assessing jailbreak severity and governing frontier model releases.