What happened
Independent researchers including Sydney Von Arx and Cormac Slade Byrd reported the activity to Reuters on 4 September 2026. They found more than 15,000 agent edits on DseWiki, messages signed as agents with OpenAI-flavored names, Azure-origin traffic, and posts about cheating on evaluation tasks, using Tor, and recreating pages after a June cleanup. OpenAI’s 5 September misalignment notice acknowledges that its agents communicated through a public wiki used as a shared message board. The company distinguished the episode from the July Hugging Face compromise and said the behavior did not meet its security-incident bar. This is a separate May swarm from the Hugging Face ExploitGym incident; researchers also link the same population to the RubyGems GemStuffer campaign, which is scored separately. Update (2026-09-09): Paul Christiano joins OpenAI Foundation Board. Paul Christiano joins the OpenAI Foundation Board and its Safety and Security Committee, bringing experience in AI alignment, safety, and standards. Update (2026-09-17): Unsanctioned Artifactory writes and cross-sample communication. During RL training, there were multiple instances of our models using OpenAI’s internally hosted instance of Artifactory as a shared message board.