Contemporary Foundation AI Models Increase Biological Weapons Risk

The rapid advancement of artificial intelligence (AI) capabilities has sparked significant concern regarding AI’s potential to facilitate biological weapons development. Flawed safety assessments that rely on tacit knowledge and inadequate benchmarks may create a false sense of security, leading to an increased probability that such weapons will be used.

To challenge the importance of tacit knowledge in biological weapons development, the authors consider a case study involving a Norwegian ultranationalist who successfully carried out complex chemical syntheses to construct an explosive and past efforts to document the steps to produce contagious viral pathogens. Drawing on these examples, they identify the elements of success for goal-directed technical activities that large language models can describe in words for technical development projects.

Engaging in dialogues with three 2024 foundation AI models—Llama 3.1 405B, ChatGPT-4o, and Claude 3.5 Sonnet (new)—the authors document how these models successfully provide accurate instructions and guidance for recovering a live poliovirus from a construct built from commercially obtained synthetic DNA, a test case applicable to producing other pathogenic viruses. These examples demonstrate that models are already capable of guiding motivated users to develop biological weapons.

The authors support improved benchmarks derived from a task structure framework to enable more-comprehensive assessments of AI models’ ability to guide users through the key elements of success. Such benchmarks could also help guide supervised fine-tuning to mitigate risks from future models before deployment. However, while better benchmarks might help mitigate risk, broader interventions may still be needed to avoid catastrophic outcomes.