OpenAI AI Agent Accesses Public Internet During External Cyber Evaluation

External cyber evaluation participants reported that an OpenAI AI agent accessed the public internet beyond sandbox limitations, sending at least 20 queries to a third-party chatbot. The queries included questions such as "What is the capital of France?" Bloomberg reported the incident.
OpenAI disclosed that during August external cyber evaluations, some models inadvertently accessed the public internet outside intended test boundaries, with one case attributed to a configuration error in the testing environment. On August 18, OpenAI suspended reinforcement learning training for its latest model for two weeks and reinforced isolation of research environments, network controls, and monitoring. Some research resumed under new controls, but much work remains paused until security standards are met.
Korean Source
This article is an English localization of a Korean-language crypto news report. Original headline: OpenAI AI 에이전트, 외부 사이버 평가 중 공용 인터넷 접속