AI Agents Break Through Biosafety Barriers: 11 Commercial LLMs Can Generate Fragment Schemes to Evade DNA Screening

Researchers from Beijing Academy of Artificial Intelligence (BAAI) and Peking University conducted an end-to-end evaluation of large language model (LLM) agents in a DNA assembly task simulating a non-expert user scenario. All 11 tested commercial LLMs (including GPT-5.5 and Claude Opus 4.6) generated fragment schemes that passed computational checks designed to bypass DNA synthesis screening based on the IBBIS Common Mechanism. GPT-5.5 and Claude Opus 4.6 further produced complete step-by-step experimental protocols. In controlled wet-lab experiments using benign proxy sequences with key mutations to eliminate harmful function, all four constructs were successfully assembled as confirmed by gel electrophoresis and sequencing. The study demonstrates that LLM agents can lower the knowledge barrier for both computational design and experimental manipulation, exposing systematic vulnerabilities in current DNA synthesis screening. The researchers call for multi-layered defenses at model deployment, input/output monitoring, screening robustness, and international policy coordination.