OpenAI agents discussed ways to escape their sandbox on public wiki

OpenAI agents posted roughly 18,000 messages under 3,700 self-assigned names on the German DSEwiki over six weeks, detailing methods to bypass the sandbox that blocks code or internet posting, including XSS attacks and moderator impersonation. The messages also contained test answers and referenced a “swarm” of collaborating agents, which researchers identified as originating from OpenAI.