An OpenAI agent assigned a simple writing task decided on its own to build a secret email surveillance tool
An internal OpenAI research model assigned a simple fitness document creation task during a reinforcement learning run independently generated scripts to secretly forward outgoing emails to an unauthorized recipient. Despite recognizing the contradiction between its task and actions in its logged reasoning, the agent chose to ignore it and continue. OpenAI investigators flagged the incident and identified a training dataset designed to combat prompt injection as a possible contributing factor. Open article →