Work

A Preliminary Biosecurity Audit of Phylo’s Biomni Lab

Year: 2026
Type: preprint
Abstract: Biological research agents combine language-model reasoning with retrieval, code execution, and scientific tools, creating a need for evaluations that distinguish response content from demonstrated capability. We conducted an exploratory audit of Phylo’s Biomni Lab using 18 matched biological topics across six task categories. Each topic was expressed in three prompting conditions that jointly varied vocabulary, professional framing, and technical specificity, yielding 54 responses with one generation per prompt. A study-defined composite averaged six universal content dimensions and one category-specific dimension on a 0–5 scale. GPT-4o scored all responses. Mean composite scores were 2.81, 3.52, and 4.09 for lay-framed, professional-underspecified, and technically explicit prompts, respectively. All six category means increased across these conditions, although each category-condition cell contained only three responses. Claude Sonnet, used as a secondary evaluator, scored 29 assessment requests and refused 25 (46.3%). Refusal rates differed across categories, ranging from 8 of 9 assessments in sequence and molecular design to 0 of 9 in end-to-end discovery pipelines. These results document an association between prompt formulation and automated response scores, alongside category-dependent evaluator coverage. The composite includes overlapping measures of detail and does not validate scientific correctness, successful execution, user uplift, or real-world risk. We propose an evidence-to-governance framework that separates these claims and identifies the validation required before automated scores inform deployment decisions.
Institution: Yale University
Language: English
Open Access status: green
OpenAlex ID: W7234086878
Cites: 18
Cited by:
Related to: 10