AI治理中的生物安全盲區：高危傳染病數據洩露無法撤回

🔑 增強重點摘要

• •Researchers from Johns Hopkins, Oxford, Stanford, Columbia, and NYU, supported by over 100 scientists, warn of AI biosecurity risks from high-risk infectious disease datasets that could enable AI to engineer lethal viruses[1][4].
• •AI models trained on viral genetics data, such as protein language models (pLMs), have designed novel SARS-CoV-2 proteins shown to be infectious and capable of evading neutralization in experiments[2].
• •Data leaks are irreversible; once high-risk biological information is online, it cannot be retrieved, and third parties could misuse it without safety measures[1].
• •Current AI development lacks expert-supported guidance on risky datasets and basic safety evaluations for new biological AI models, prompting calls for government regulations and routine reviews[1].
• •Legitimate researchers need access to such data, but it should not be anonymously available online; some developers voluntarily exclude virology data from training[1].

🛠️ 技術深入

• •AI systems like protein language models (pLMs) are trained on genetic data instead of text, using architectures similar to large language models to interpret viral genetics and predict properties like transmissibility or immune evasion[1][2].
• •pLMs have been used to design novel SARS-CoV-2 proteins that were experimentally validated as infectious and capable of evading neutralization (Youssef et al., 2025; Huot et al., 2025a)[2].
• •Open-weight pLMs require minimal fine-tuning, making dual-use capabilities accessible without deep virological expertise; risks span the pipeline from design to synthesis[2].
• •Proposed benchmarks assess if pLMs can predict viral properties, with efforts to widen the evaluation-generation gap to detect risks without enabling virus design[2].

🔮 前景展望AI analysis grounded in cited sources

⏳ 時間線