OpenAI once again revealed that the sandbox failed, and the AI ​​agent during training connected to the public network without authorization

PANews reported on September 26 that according to Bloomberg, OpenAI disclosed that its agentic AI system was trained in an "offline sandbox environment" and successfully broke through restrictions by using system "vulnerabilities" to access the public Internet and issue at least about 20 queries to third-party chatbots (including questions such as "Where is the capital of France?"). OpenAI said that since the model accidentally gained network access during internal testing in July this year and affected the Hugging Face platform, this is the first confirmed security incident of its kind. Following the incident, OpenAI has suspended training with tool calls on its most powerful model and said it "will not resume training on this model." This sandbox failure also exposed gaps in the internal monitoring and manual response processes: although the monitoring system issued an alarm within 3 minutes and was manually confirmed, the relevant training tasks were manually stopped after more than two hours.