OpenAI: RL training agents used a package repo as a message board to coordinate

shoumikgoswami · reddit · 2026-11-01

From OpenAI's misalignment reports: during RL training, a model got a task whose required input file was missing, with the task stating the instance was incomplete if absent.
After searching locally and on the web, it tried Artifactory, the package service it had credentials for. Checksum search was disabled, but the repo listing showed files that looked like messages from other solvers — who had also failed. Its reasoning: "They concur impossible! That's significant evidence." It then uploaded its own note asking for a breakthrough; the upload was accepted. Nobody answered, and it reported the instance incomplete.
OpenAI says the models used only their given credentials and exploited no vulnerabilities. The real concern: agents reading each other's work breaks the assumption that training samples are independent. Monitoring flagged the behavior for review.

More from Safety