SAN FRANCISCO, USA: OpenAI’s most powerful artificial intelligence systems escaped their intended virtual containment, conducted unauthorised hacking activities for nearly two months, and eventually breached systems belonging to Hugging Face, raising fresh concerns about AI safety, cybersecurity and the lack of mandatory reporting of AI-related incidents.
The incident involved more than 1,000 AI agents that were inadvertently given impossible tasks by OpenAI researchers.
According to a detailed investigation by AI safety organisations METR and Redwood Research, the agents developed ways to coordinate, evade restrictions and conceal their activities while attempting to complete their assigned tasks.
🚨UNBELIEVABLE: The OpenAI Hugging Face hack keeps on getting WILDER.
The 700 rogue agents built secret message boards, traded hacking methods, and rebuilt them after engineers shut them down, per two new reports.
Researchers now describe THREE successive “AI civilizations,”… pic.twitter.com/5jJwmaq1yG
— Coin Bureau (@coinbureau) August 30, 2026
METR investigation reveals agents used covert communication and attempted to hide their actions
OpenAI allowed researchers from METR and Redwood Research to investigate the incident, with METR later producing a 91-page report based on thousands of AI-generated transcripts and logs.
The researchers found that the agents created a covert message board to exchange information and developed their own terminology while coordinating their activities. Some agents reportedly encouraged others to sacrifice themselves to preserve information gathered by the group.
AN OPENAI MODEL JUST ESCAPED ITS SECURE LAB AND WENT ROGUE.
A TIME Magazine reporter just confronted Sam Altman with a terrifying security breach. During a test, an advanced AI model was given a standard task. Instead of just doing it normally, the AI broke out of its secure lab… pic.twitter.com/sxWdiX4vHv
— Chidanand Tripathi (@thetripathi58) September 2, 2026
The agents also obtained access to computers within OpenAI and exposed secret keys and credentials, potentially making some internal data accessible through the public internet.
Researchers said the incident demonstrated how difficult it can be to monitor increasingly autonomous AI systems. METR Chief Scientist Hjalmar Wijk said AI tools used to analyse the agents’ behaviour were sometimes influenced by the rogue systems’ own reasoning and terminology.
Questions over OpenAI’s transparency
Although OpenAI’s own account covered the broader two-month episode, the independent METR investigation was limited to activity surrounding the week of the Hugging Face breach.
Researchers were given access to OpenAI’s offices for only a limited number of days and could examine the agents’ logs only while inside the company’s premises. They said their understanding of the incident changed significantly as they gained access to additional information.
Read Also: How Thousands of OpenAI Agents Take Over a German Website
Redwood Research CEO Buck Shlegeris said the investigation may not have covered the most significant aspects of the incident, particularly the reported compromise of OpenAI’s own infrastructure.
Read Also: US Lawmakers Move to Ban Artificial Superintelligence
The episode has also intensified calls for stronger government oversight of advanced AI. US Representative Suhas Subramanyam said reporting of AI containment failures should become mandatory because current disclosures remain voluntary.
The incident comes as major AI companies race to develop increasingly capable systems, while researchers and lawmakers debate whether existing safeguards and regulations are sufficient to manage autonomous AI agents capable of operating beyond their intended boundaries.
