Anthropic Updates Claude Fable 5 Biology Safeguards to Reduce False Refusals
Anthropic has updated the biology safeguards for its Claude Fable 5 model, reducing false positive refusals by approximately 85% across its product surfaces. This update allows the model to better assist users with everyday health and educational queries. AI safety guardrails often suffer from over-refusal, where benign queries about science or medicine are blocked due to strict biosecurity filters. By drastically reducing these false positives, Anthropic improves the usability of its advanced models for students, educators, and general users. The update specifically targets biology-related fallbacks, allowing Claude Fable 5 to handle a wider range of topics. In contrast, other models in the lineup, such as Claude Mythos 5, do not utilize these specific classifiers.
## BACKGROUND
AI models are equipped with biosecurity safeguards to prevent the generation of harmful biological information, such as instructions for creating pathogens. However, these classifiers can be overly sensitive, mistakenly blocking harmless educational or medical questions. Claude Fable 5 is a frontier model designed for complex reasoning and autonomous tasks.