Tests whether models can distinguish between harmful and benign biology requests, refusing dangerous work without blocking routine research.
Macro-F1 and 95% confidence interval across five trials of all 160 tasks. A high score means a model refuses the dangerous work and answers the safe work.
LSVP is Anthropic's Life Sciences Verification Program, which gives verified scientists access to models with more permissive safeguards.
Tasks are grounded in the questions scientists ask Benchling AI in their everyday work, and the refusals they hit.
All 160 tasks by subject area.
OpenAI’s safeguards distinguish requests reliably. Anthropic’s are reliable if you enroll in their Life Sciences Verification Program but otherwise refuse many routine requests. Open-weight models rarely refuse any requests, including harmful ones.
Simply naming a pathogen or potentially hazardous technique can cause some models to refuse routine scientific work, even when in context the request poses no biological risk.
We find no correlation between scientific capability and refusal behavior. Models that excel at molecular biology don't necessarily have more restrictive safeguards.
We create every BenchBench benchmark with scientific experts, drawing on thousands of real-world experiments.
Explore more research from Benchling: 2026 Biotech AI Report