Anthropic's Claude Fable 5 blocks basic biology queries to prevent bioweapon misuse
The new Mythos-class AI refuses to answer 'what are mitochondria' or explain cell membranes.
Anthropic released Claude Fable 5, its first Mythos-class model, calling it the most capable AI it has ever made widely available. However, The Verge reports that the model will not answer basic biology questions—such as 'what are mitochondria' or 'explain how antibiotic resistance arises'—because of intentionally over-conservative safety guardrails. Anthropic told The Verge that the filters are designed to block queries tied to bioweapons research, and that this tradeoff allows customers to use the model's advanced capabilities sooner without the risks. When Fable refuses a question, it typically hands off the query to the former flagship Claude Opus 4.8, which answers freely. While biology is the most restricted domain, the model also limits certain chemistry and cybersecurity queries, though it answered basic questions about TNT and chlorine gas. Anthropic admits the current safeguards produce false positives and is working to improve detection precision. The company has previously flagged chemistry, biology, cybersecurity, and distillation as key areas where it would throttle Fable's responses for safety.
- Claude Fable 5 refuses to answer harmless biology queries like 'what is a prion' or 'how do mRNA vaccines work', deferring to Claude Opus 4.8.
- Anthropic says the overly conservative filters are a deliberate safety tradeoff to prevent bioweapon misuse while still releasing the powerful Mythos-class model.
- The model is more permissive on chemistry and cybersecurity (e.g., explained TNT basics and chlorine gas), but blocks synthesis instructions and highly toxic agents like sarin.
Why It Matters
Shows the tension between releasing powerful AI and preventing misuse, forcing hard tradeoffs that can limit even benign use.