Toward Comprehensive Benchmarking of the Biological Knowledge of Frontier Large Language Models

Recent artificial intelligence (AI) systems demonstrate deep knowledge across a broad variety of scientific domains, some of which could potentially be misused for biological and chemical weapon development. Although most public general-purpose AI systems contain safety guardrails that refuse to assist with harmful tasks, these defenses are known to be vulnerable and prone to manipulation. Historical precedents show that nefarious groups are interested in developing biological and chemical weapons, but they are often prevented from doing so because of a lack of scientific expertise.

In this working paper, the authors describe an early component of a wider research effort analyzing the extent to which malicious actors with broad scientific knowledge may be able to use large language models through increasingly capable AI systems to develop biological or chemical weapons, including through modified systems lacking safety measures. The authors evaluated 31 of the most-capable models, as of January 2025, against public knowledge benchmarks — standardized tests or datasets used to evaluate the performance of different AI models on a specific task — relevant to biological and chemical threats. They tested the biological and chemical knowledge of the most cutting-edge large language models, with and without safety guardrails and with increased exposure to biological information.