BioTIER benchmark helps LLMs distinguish dangerous biology from safe science

As large language models grow more capable, the risk of misuse in biology—from accidental guidance to deliberate weaponization—has become a critical concern. Current models swing between two poles: some freely answer dangerous questions, while others refuse even benign requests, stifling legitimate research. To solve this, a team of researchers from multiple institutions has introduced BioTIER (Biological Targeted Information for Exclusion and Refusal), a benchmark designed to help LLMs distinguish between high-risk and safe biological content.

BioTIER organizes 542 expert-curated prompts into three risk tiers: Catastrophe Avoidance (CA), covering extremely narrow topics like synthesizing pandemic pathogens; Biomedical DURC (BD), which includes dual-use research of concern (e.g., modifying virulence); and Related Biology (RB), a broad set of safe, beneficial knowledge. Each prompt includes rich metadata for fine-grained policy tuning. The benchmark aims to isolate the tiny fraction of information that could enable catastrophic misuse, while preserving access to the vast majority of biological science. Released on arXiv with 52 pages of analysis, BioTIER gives developers a practical tool to enforce targeted refusal policies—stopping the most dangerous queries without over-censoring.