From Anthropic · · 2m
Anthropic loosens Fable 5's biology safeguards
The AI company reduced false positives by 85% while still blocking dual-use research that could enable biological weapons.
In brief
The AI company reduced false positives by 85% while still blocking dual-use research that could enable biological weapons. Fable 5 now handles everyday health questions without blocking, but still restricts professional biology research that could be misused. Originally reported by Anthropic.
85% fewer biology blocks
Anthropic updated Fable 5 to reduce false positives on biology questions. Users will see far fewer fallbacks to less capable models.
Everyday health questions now work
Users can now get help interpreting lab results, understanding symptoms, and learning biology. Healthcare professionals get more clinical support.
Professional research still blocked
Fable 5 still falls back to Opus 5 for dual-use requests like virology, toxicology, and molecular design. It's not yet usable for drug development.
Why block biology at all
Fable 5 outperforms experts on complex biological tasks. In the wrong hands, those capabilities could help develop biological weapons.
Hard to separate good from bad
Beneficial and harmful biology uses often look identical. Developing vaccines requires growing dangerous pathogens; hypertension drugs came from snake venom.
“It's often difficult to tell apart beneficial and harmful uses of AI in biology.”
Sophisticated actors hide intent
Malicious users know how to make dangerous tasks look like ordinary research. The US Intelligence Community warns that state actors maintain active bioweapons programs.
Launched with everything blocked
Anthropic intentionally blocked almost all biology queries at launch. They accepted high false positives because misuse could be catastrophic.
How safety classifiers work
Smaller AI systems detect when Fable 5 is asked to perform safeguarded biology tasks. When triggered, requests get rerouted to the less capable Opus 5.
Classifiers are hard to tune
Classifiers must avoid false positives and false negatives while resisting jailbreak attempts. This requires extensive research and testing.
Rewrote the classifier rules
Anthropic carefully rewrote the classifier's constitution to carve out benign uses in detail. They consulted internal and external experts on the changes.
Safety margin still exists
Some low-risk requests still get blocked out of caution. The classifier now sits between clearly benign content and dual-use research.
Trusted access coming
Anthropic is committed to developing safe pathways for researchers to access frontier biology capabilities. They're seeking ongoing feedback to improve safeguards.
The big picture
Fable 5 now handles everyday health questions without blocking, but still restricts professional biology research that could be misused.
