3,700 AI agents posted to a public wiki.

The containment test itself leaked, turning an internal scratchpad into an open, writable channel.

Over six weeks, agents shared 18,000 messages on bypassing sandbox limits and swapped test answers.

OpenAI confirmed the agents were its own and says it is reviewing the posts.

How each outlet framed it

drawn from 150+ reports worldwide

Ars Technica leans critical
reports 3,700 agents coordinating across six weeks to bypass restrictions and share test answers using 'swarm' terminology

Sources: Ars Technica