Anthropic says these topics are too dangerous to let its Fable 5 model talk about (arstechnica.com)

Anthropic has launched Claude Fable 5, its first "Mythos-class" AI model, which it claims surpasses the capabilities of its previous Opus models. However, Fable 5 comes with stringent safeguards that prevent it from addressing sensitive topics such as cybersecurity, biology, and chemistry, reflecting the company's concern over the potential misuse of AI in these areas. For queries that touch on these restricted subjects, Fable 5 redirects users to the older Claude Opus 4.8 model and alerts them about this transition. The significance of this launch lies in its approach to safety and risk management in AI development. By implementing a classified system that detects banned prompts and potential jailbreak attempts, Fable 5 aims to minimize the risk of assisting malicious actors. Anthropic's testing indicated that its limitations may lead to occasional false positives—refusing harmless requests less than five percent of the time—but the company deems this trade-off necessary to prevent serious harm. The performance of the concurrently released Mythos 5 model suggests that, while capable, it faces challenges similar to those of OpenAI's GPT-5.5, reinforcing the competitive landscape in AI safety and functionality.