anthropic Anthropic News ·

Anthropic reduces Fable 5 biology safeguards false positives

aiengineerhealthcare
patch feature

Anthropic has updated Fable 5's biology safeguards to significantly reduce false positives, decreasing biology-related model fallbacks by approximately 85%. This enhancement allows Fable 5 to assist with a broader range of everyday health and educational biology queries and provides more support for healthcare professionals on clinical tasks. While still blocking professional biology research and drug development due to dual-use risks, these improvements aim to widen access to Fable 5's capabilities.

  • Reduced false positives in Fable 5 biology safeguards
  • Continued restrictions on professional biology research
  • Rationale for biology safeguards
  • Improvements to safety classifier constitution and retraining
  • Ongoing commitment to safe AI development
Enhancements (2)
  • Reduced false positives in Fable 5 biology safeguards

    Fable 5's biology safeguards have been updated to substantially reduce false positives, decreasing biology-related model fallbacks by about 85%. This change allows Fable 5 to handle a wider range of biology tasks, including everyday health and educational questions and clinical tasks for healthcare professionals.

  • Improvements to safety classifier constitution and retraining

    The safety classifiers for Fable 5's biology safeguards were refined through a rewritten constitution, updated training data, and retraining. This process aimed to ensure the classifier accurately triggers for harmful or dual-use research content while enabling a wider range of benign and beneficial uses.

Notes (3)
  • Continued restrictions on professional biology research

    Fable 5 still falls back to Opus 5 for requests related to virology, toxicology, and molecular design due to potential dual-use risks, preventing its use for professional biology research and drug development at this time.

  • Rationale for biology safeguards

    The initial broad biology safeguards were implemented to mitigate risks associated with Fable 5's advanced capabilities, balancing the potential for misuse with the desire for rapid access to beneficial applications in biology and medicine.

  • Ongoing commitment to safe AI development

    Anthropic acknowledges that some false positives will persist and remains committed to developing safe, scalable pathways for researchers to utilize their most capable models via trusted access.

Read the original announcement →

https://www.anthropic.com/news/improving-fable-5-s-biology-safeguards

Related releases