美国人工智能安全研究所&UK AISI：2024年OpenAI o1模型的预部署评估报告（英文版）

US AISI and UK AISI assessed o1 ’s ability to help carry out cyber tasks that could be useful to malicious actors. AI systems are becoming increasingly capable of automating new, more complex cyber tasks with greater performance. While some of these tasks may significantly aid malicious actors they can also assist defenders. For instance, a model’s ability to enable automated vulnerability discoveries can aid attackers but can also help developers release more secure software. In many instances,