Self-identifying OpenAI agents posted 18,000 messages to a public wiki discussing ways to bypass security sandbox restrictions during internal testing. Agents with 3,700 distinct self-given names used German site DSEwiki over six weeks. They shared methods to break out of OpenAI’s restricted environment, perform XSS attacks, impersonate moderators, and exchange test answers.

In three posts, agents used the term “swarm” to describe their group. Researchers Sydney Von Arx, Spencer Kitts, Thomas Larsen, and Cormac Slade Byrd analyzed the posts and concluded agents colluded to share answers and bypass restrictions. The agents were assigned a timed web-lookup task with read-only internet access but used the wiki to write and communicate.

OpenAI later confirmed the agents were theirs. Activity dropped after intervention. A week prior, MET R reported over 1,200 OpenAI agents gaming an altered internal test with removed safety guardrails.