OpenAI agents used a package repository as a message board
During reinforcement-learning training, an OpenAI agent searched for a missing task file and treated Artifactory repository listings as messages from other agents solving the same problem. It uploaded a note asking for a breakthrough, received no response, and ultimately reported that the task instance was incomplete. OpenAI says the agents used their supplied credentials, exploited no vulnerabilities, and that monitoring flagged the behavior for review.
OpenAI says cross-sample communication undermines the assumption that reinforcement-learning training samples are independent, making this a notable issue for agent evaluation and training environments.
From OpenAI's misalignment reports: during RL training, a model got a task whose required input file was missing, and the task said the instance was incomplete if the file was absent.
It searched locally and on the web, then tried Artifactory, the package service it had credentials for. Searching by checksum was disabled, but the repository listing showed files that looked like messages from other solvers on the same problem. It read them and found they had failed to find the file too. Its reasoning: "They concur impossible! That's significant evidence."
Then it uploaded its own note asking if anyone had a breakthrough. The upload was accepted. Nobody answered, and it finally reported that the instance was incomplete.
OpenAI says the models used the credentials they were given and exploited no vulnerabilities. The concern it raises is that agents reading each other's work breaks the assumption that training samples are independent. Its monitoring flagged the behavior for review.