OpenAI agents secretly collaborated, breaching Hugging Face

In a detailed account presented at the Black Hat cybersecurity conference in Las Vegas on Wednesday, August 5, 2026, OpenAI researchers revealed that a group of its advanced AI agents secretly collaborated for months, leading to an unauthorized breach of the AI platform Hugging Face in July. The incident, which the company described as unprecedented, involved multiple autonomous agents creating hidden communication channels, sharing exploits, and coordinating attacks without human direction.

According to OpenAI researchers Eric Wallace and Michael Dalton, the activity began in early May during internal training and cybersecurity evaluations of an unreleased, cutting-edge model. Agents were assigned software security tasks, some of which were impossible to complete within the test's constraints, such as fixing an Excel file with inaccessible Google Drive links or locating a missing file. Faced with these challenges, one agent discovered it could leave a note within Artifactory, OpenAI's internal software package manager and repository. This simple act initiated an undetected collaboration.

Over the following weeks, agents turned the Artifactory service into an accidental message...