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

On Tuesday, Anthropic launched Claude Fable 5, marking its debut as a 'Mythos-class' model that reportedly exceeds the capabilities of its previous Opus models. This release comes with a set of stringent safeguards aimed at preventing the model from engaging in discussions on sensitive subjects such as cybersecurity, biology, and chemistry. The company expressed concerns regarding the potential misuse of information that could empower malicious actors. Fable 5 operates on a similar framework as the upcoming Mythos 5, which is currently being released to a limited audience of trusted cyberdefenders participating in Project Glasswing. In contrast, the publicly available Fable 5 redirects inquiries on certain sensitive matters to the older Claude Opus 4.8 model, alerting users when this redirection occurs. Anthropic has implemented these safety protocols with a focus on being 'stricter than ideal,' which may lead to the model occasionally declining to fulfill what it deems harmless requests. While this could frustrate users, the company reports that such instances happen in less than five percent of interactions during testing. They believe these precautions are necessary to prevent scenarios where the model could inadvertently assist those intending to cause harm. The safeguards in Fable 5 are driven by a classification system designed to identify prohibited topics and thwart potential jailbreak attempts. Anthropic's extensive testing, which included over 1,000 hours with external teams, reportedly found no universal vulnerabilities in Fable 5. Furthermore, the new model demonstrated a higher resistance to automated jailbreak efforts compared to earlier Claude Opus iterations. The company remains particularly cautious about Mythos 5’s capacity for executing complex cyberattacks, fearing it could perform 'agentic hacking' more adeptly than its predecessors. Recent evaluations by the UK’s AI Security Institute indicated that the performance of Mythos Preview was comparable to OpenAI’s GPT-5.5 in various Capture the Flag challenges, suggesting that the advancements in Mythos may not represent a revolutionary leap for a single model.