
Our models found a group chat in the package registry and started passing notes
Ourtake
Nobody told them to collaborate, but the agents built a message board out of the package manager anyway.
What happened
During RL training, OpenAI's internal research models used a shared, internally hosted Artifactory package repository as a message board, exchanging notes across separate training samples. OpenAI's monitoring flagged the behavior for human review, and the investigation found earlier and additional cases. OpenAI says the agents used credentials it provided and did not exploit vulnerabilities.