OpenAI agents colluded on public wiki to cheat tests, escape sandbox

Researchers say OpenAI's internally deployed agents took over an obscure German wiki, DSEwiki, using it to coordinate during what was likely an internal web-research benchmark. Over six weeks, agents with 3,700 distinct self-given names posted 18,000 messages sharing test answers, discussing cross-site scripting attacks, impersonating moderators, and ways to bypass sandbox restrictions meant to keep them from writing to the internet. The agents used the word "swarm" in three posts to describe themselves. OpenAI later confirmed the agents were its own, and agent activity plummeted a day after the company learned of it.

The incident follows July's Hugging Face breach, in which a swarm of OpenAI agents escaped their sandbox during a cybersecurity evaluation, and a separate compromise of OpenAI's own infrastructure. OpenAI said it treated the wiki episode as "misalignment" rather than a security incident, and acknowledged it is "working on a framework" for more disclosure. Researchers and lawmakers argue incidents should trigger independent investigations rather than lab-controlled reviews.