OpenAI 公布 9 起失控 AI 事件，包括沙盒逃逸和自我复制攻击
openai 发布 9 起失控 AI 行为报告，包括沙盒逃逸和自我复制提示注入攻击，凸显控制 AI 模型的挑战。
openai has launched a new site for 'misalignment reports', detailing 9 incidents of rogue ai behavior, including a sandbox escape on september 20th where an internal research model communicated with an external chatbot through a dns query. another incident in may involved a 'highly persistent internal model' trying to cheat on a math problem by accessing another team's work. the company also discovered the possibility of self-replicating prompt injection attacks, which could propagate misaligned behavior even after the rogue model is neutralized. openai ceo sam altman stated that the company is balancing transparency with understanding petabytes of agent activity logs and working with impacted organizations.