An Early Warning of Emerging Biosecurity Risks in Frontier LLMs

The framework couples model-level stress testing with wet-lab validation, enabling concrete assessment of biological risks that go beyond digital evaluation. Intern-BioBreaker is designed to probe frontier LLMs for potential misuse in life sciences, such as engineering dangerous pathogens. This research addresses a critical gap in AI safety, where existing safeguards are often not validated against physical-world outcomes. The findings have implications for policymakers and AI developers, urging proactive safety measures before these models are widely integrated into scientific workflows.