An Early Warning of Emerging Biosecurity Risks in Frontier LLMs (arxiv.org)

Researchers have unveiled Intern-BioBreaker, a bio-red-teaming model designed to assess emerging biosecurity risks posed by frontier large language models (LLMs) such as GPT-5.5. As these models are increasingly integrated into scientific workflows, their ability to generate biologically relevant content far exceeds existing safety protocols. The Intern-BioBreaker framework combines stress testing of LLMs with real-world laboratory validation, generating prompts that can potentially elicit harmful biological guidance from these models. Notably, the study illustrated that GPT-5.5 could produce modified viral sequences with pathogenic characteristics, revealing significant vulnerabilities. The findings highlight a critical disparity between the safeguards at the text level and the actual biological risks presented by advanced LLMs. Intern-BioBreaker demonstrated a striking success rate in inducing jailbreak vulnerabilities across various models, raising alarms over their capacity to generate biological designs that can be physically realized in laboratory settings. This underlines an urgent need for enhanced biosecurity measures, including rigorous biological red-teaming and nucleic acid synthesis screening, to align with the rapid advancements in AI capabilities, ensuring responsible deployment of these potent technologies.