Anthropic says these topics are too dangerous to let its Fable 5 model talk about

Anthropic released Claude Fable 5 on Tuesday, its first "Mythos-class" model which it claims surpasses its previous Opus models in overall capabilities. However, Fable 5 includes safeguards designed to prevent it from responding to queries on sensitive topics such as cybersecurity, biology, and chemistry, due to concerns about potential misuse. The company stated that Fable 5 operates on the same underlying model as Mythos 5, which is concluding its "Mythos Preview" period today for a select group of cyberdefenders vetted through Project Glasswing. Unlike Mythos 5, the publicly available Fable 5 is engineered to redirect queries on certain sensitive subjects to the earlier Claude Opus 4.8 model and to notify users when this redirection occurs. Anthropic reported significant benchmark improvements for Fable 5, with a particularly notable advancement in cybersecurity-related performance. The company acknowledged that these safeguards are tuned to be "stricter than ideal," potentially leading to refusals of harmless requests in less than five percent of testing sessions. Anthropic justified this approach by emphasizing the importance of preventing malicious actors from obtaining assistance that could lead to serious harm, which they might not be able to acquire from other sources.