Anthropic Releases Guarded Claude Fable 5 Model

· Original article ↗

Summary

Anthropic says Claude Fable 5 routes cybersecurity and biology queries to Claude Opus 4.8, adds safeguards, and passed testing that found no known universal jailbreaks. The company acknowledges some benign requests may be blocked.

Key points

  • Anthropic says Fable 5 is based on Claude Mythos but routes queries on a narrow set of cybersecurity and biology topics to Claude Opus 4.8.
  • The company says internal and external red-team tests found no known universal jailbreaking techniques; the article notes Anthropic did not say whether partial techniques were found.
  • Anthropic cautions that safeguards may block some benign requests and says it plans to refine them after launch.
  • Anthropic says Fable and Mythos user traffic will be retained for 30 days and not used to train new Claude models or for non-safety purposes.
  • Project Glasswing members are to receive access to Claude Mythos 5 through a trusted-access program, with expansion planned over time.
  • The article cites Opus 4.8 system-card tests: safeguards reportedly reduced reproduction of previously discovered open-source vulnerabilities from nearly 80% to 1%.

Article Details

Event Type
Public release of Claude Fable 5, a guarded version of Claude Mythos.
Impact
Anthropic says cybersecurity and biology queries are routed to Claude Opus 4.8 instead of the full model to reduce misuse risks. The company also says safeguards may block some benign requests and that user traffic for Fable and Mythos will be retained for 30 days, not used to train new Claude models or for non-safety purposes.

Vendors

Products

Related Articles