Biorisk
Also called Biological Risk
The danger that a model meaningfully helps someone create a biological weapon.
Think of it like
Worrying less about the library having a chemistry book and more about it having a personal tutor who walks you through every step.
Example
Red-teamers test whether a model gives non-experts actionable uplift toward synthesizing a dangerous pathogen, and safety training is tuned to refuse it.
How it actually works
The concern is "uplift" — not whether facts exist online, but whether the model lowers the skill barrier by planning, troubleshooting, and connecting steps. It is among the most tightly guarded categories, with dedicated classifiers, refusals, and expert-run evals. Measuring it responsibly is itself sensitive work.
For product teams
One of the highest-severity risks, often subject to explicit safety commitments and external testing.
For engineers
Evaluate end-to-end task uplift for non-experts, not fact recall; guard with layered refusals and dedicated classifiers.
Related
- Dangerous capabilities — A headline category within dangerous capabilities.
- Cyber risk — Sits alongside cyber risk as a top frontier concern.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome