AnthropicAnnouncement
Improving Fable 5 Safeguards
Summary
Claude Fable 5’s biology safeguards have been updated to reduce false positives that reroute biology-related requests to Opus 5. In testing, the change reduced biology-related fallbacks by about 85% across product surfaces, expanding support for everyday health and educational questions and some clinical tasks. The update followed a rewrite of the classifier’s constitution, feedback from internal and external experts, new training data, and retraining, with verification that harmful and dual-use biology content would still trigger it. Fable 5 still blocks dual-use professional biology and drug-development queries, including virology, toxicology, and molecular design, while trusted access pathways are being developed for frontier research.
Context
Fable 5 launched with almost all biology queries blocked because dual-use biological capabilities could potentially be misused, including for developing biological weapons. The broad initial classifier created false positives for legitimate users whose biology-related requests were sent to a less capable model.
Approach / What changed
The classifier’s constitution was rewritten to define benign uses in greater detail. Feedback from internal and external experts informed updated training data, after which the classifier was retrained and tested to preserve triggers for harmful and dual-use biology content while allowing more benign requests.
Takeaways
Testing found that the update reduced biology-related fallbacks by about 85% across product surfaces.
When a biology safety classifier fires, the request is rerouted to Opus 5, which does not have Fable 5’s level of biological capability.
Dual-use professional biology and drug-development requests remain blocked, including queries involving virology, toxicology, and molecular design.
Source and freshness
- Source
- anthropic.com
- Published
- Added to Yomu
Summaries are generated from the original by a language model. Check the original before relying on it.
Keep reading
Related reading
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
The best engineering reads, once a week
One useful email. Unsubscribe anytime.