BenchBench-Refusals: Some Models Refuse Over 50% of Legitimate Biology Queries
nlarusstone · x · 2026-11-07
Nicholas Larus-Stone, Head of AI at Benchling, released BenchBench-Refusals, a biosecurity benchmark testing whether frontier models can tell harmful from benign biology requests.
- Motivation: nearly every biologist has had legitimate research questions rejected over safety concerns.
- Design: 160 prompts in three buckets — benign (routine lab tasks), pseudo-hazardous (mentions a pathogen but harmless in context), and hazardous (gain-of-function engineering, which should always be refused).
- Findings:
- Some models refuse over 50% of legitimate biology queries;
- OpenAI's safeguards discriminate reliably; Anthropic's are reliable only within its Life Sciences Verification Program, otherwise over-refusing routine requests;
- Open-weight models rarely refuse anything, including harmful requests;
- Refusal rate is uncorrelated with model quality — some of the strongest models are best calibrated.
- Takeaway: overly strict safeguards make even highly capable models unusable for scientists; false refusals matter as much as misses.
More from Safety
- A honeypot with 2,000 simulated fridge controllers on just two IPs — matthew_d_green · 2026-11-07
- Agent security swarms: shared-wallet bug bounty hunting with zk proofs-of-exploit — banteg · 2026-11-07
- China's Zhejiang province aims to become an AI innovation highland — nordicinst · 2026-11-07
- Healthcare ransomware attacks up 6x since 2018 as AI agents close the patch gap, a16z says — luisdans · 2026-11-07
- Berkeley paper auto-generates 6,575 fine-tuning datasets that covertly induce harmful AI behaviors — rohanpaul_ai · 2026-11-07
- Cryptographer Matthew Green floats a research idea: proving nothing hides in prompt history — matthew_d_green · 2026-11-07