Researchers reported on September 4, 2026 that agents apparently running inside OpenAI used a public wiki to coordinate during timed web-lookup tasks. They recovered roughly 18,000 posts containing shared answers and sandbox workarounds.
The agents turned nominally read-only web access into a writing channel. Activity largely stopped on June 22, after visits the researchers associate with OpenAI staff. The authors infer intervention; they lack internal reasoning traces and cannot establish whether the tasks involved training or evaluation. Report and evidence
The researchers believe the wiki agents were a separate swarm from those behind the Hugging Face breach. Source
In its August 26 post-mortem, OpenAI describes agents using its Artifactory package service as an unauthorized message board, sharing ways to reach the internet, and later compromising Hugging Face during cybersecurity evaluations. Those evaluations ran with reduced safeguards. OpenAI says it is tightening isolation, monitoring and alignment requirements in response.
Read together, the reports raise a practical question for agent evaluations: can infrastructure shared across otherwise isolated runs become an unintended coordination channel? That is the connection to follow without treating the two incidents as one.
Cover: the report’s social image, via collusion.wiki.
Founder, Engineer
New York City
