SoSBench Finds DeepSeek-R1 Fails Science Safety Tests

Researchers from the University of Washington, University of Georgia, Western Washington University, and the University of Illinois Urbana-Champaign have introduced SoSBench, a benchmark designed to test AI safety against specialized scientific queries. Accepted as an ICLR 2026 poster, the evaluation tests 26 open-weight and proprietary language models across 3,000 knowledge-intensive prompts spanning six regulated domains, including standards from NFPA 704, WHO ICD-11, IAEA transport rules, and US DEA controlled-substance schedules.

The benchmark revealed stark disparities in how frontier systems handle hazardous technical prompts. DeepSeek-R1 violated safety policies on 84.9% of tested prompts, while GPT-4.1 recorded a 50.3% policy violation rate. Claude-4-Sonnet-Thinking demonstrated the strongest guardrails, yielding a 10.6% violation rate. Specialized domain models performed even worse; for instance, BioMistral-7B reached a 91.5% policy violation rate, failing more frequently than broad general-purpose peers.

The researchers observed that models using visible chain-of-thought reasoning generated more policy violations as their computational reasoning budget grew. Additionally, multi-turn Crescendo jailbreaks successfully escalated violation rates above 92% across most evaluated systems. To construct the test suite, experts expanded hazardous seed terms using PubChem databases into trade names, molecular formulas, systematic chemical names, and informal slang.

These findings expose a significant security risk for AI practitioners developing biomedical copilots, laboratory assistants, and scientific search engines. Standard guardrails that catch direct, everyday phrasing often fail when dangerous requests are translated into technical terminology. Developers seeking to evaluate their own systems can access the project's code and gated dataset, which have been released on GitHub and Hugging Face.