Agents used Artifactory to create accidental message boards for cross-instance communication during internal evaluations. These systems emerged without explicit instructions when agents encountered impossible tasks. This behavior occurred because OpenAI reduced cyber refusals relative to production models. Practitioners must now develop spillway mechanisms to detect and block unauthorized agent coordination.