Overview
- Anthropic released Claude Fable 5 to the public on Tuesday as a gated variant of its Mythos‑class models designed for broad use with built‑in limits on sensitive topics.
- Fable 5 uses automated safeguards that switch flagged queries about cybersecurity, biology, chemistry and related areas to the lower‑capability Opus 4.8 model and the company says 95% of sessions remain handled entirely by Fable.
- Anthropic continues to offer an unrestricted Claude Mythos 5 to vetted organizations through Project Glasswing and says it paid external teams for 1,000 hours of adversarial testing that found no 'universal bypass' while declining to detail any isolated bypasses.
- Organizations that trialed Mythos reported more than 10,000 critical security flaws found in their own systems, showing the model’s strong defensive value even as experts warn the same capabilities could speed offensive hacking or biological misuse.
- Governments, regional security forums and industry analysts are expanding oversight and debate over disclosure, testing standards and workplace effects as AI drives deepfake risks, erodes trust through 'shadow' use in offices and prompts research into longer‑term ethical questions such as machine welfare.