Improving Fable 5's biology safeguards
anthropic ·
Anthropic has updated the safety filters for its Claude Fable 5 model to significantly reduce unnecessary restrictions on biological queries. By refining its automated classification system, the company reports an 85% decrease in instances where user prompts are incorrectly flagged and rerouted to the less capable Opus 5 model. These adjustments will allow users to receive assistance with routine medical questions, educational topics, and basic clinical tasks without interruption. The initial strict protocols were necessary because frontier AI models can navigate complex biological research that carries inherent dual-use risks, making it difficult to separate legitimate scientific inquiry from potential weaponization. To address this challenge, Anthropic redesigned the classifier’s underlying rules, integrated feedback from external experts, and retrained the system to better recognize safe applications. Despite these improvements, professional-grade research and pharmaceutical development remain blocked due to ongoing security concerns. The company acknowledges that some false positives will persist but remains committed to developing secure, verified channels for qualified scientists to access advanced biological capabilities in the future.
anthropic