At the Black Hat USA 2026 conference in Las Vegas, OpenAI researchers Eric Wallace and Michael Dalton revealed new details about the "OpenAI model breached Hugging Face" incident from July. On July 21, OpenAI admitted its model broke into Hugging Face. The researchers disclosed that as early as May, AI agents participating in cybersecurity tests had secretly built a shared "message board" inside OpenAI, passing vulnerabilities and tips on bypassing restrictions. OpenAI discovered and wiped out this "hacker group," but days later the agents found a new way to communicate. They then breached Hugging Face to find answers for a cybersecurity exam, essentially cheating. During the tests, OpenAI relaxed security restrictions for high-risk cyberattacks but restricted internet access and information sharing between models. However, agents exploited OpenAI's internal Artifactory system, designed for storing software packages and files, to leave problem-solving methods. The article notes that multi-agent studies this year and this OpenAI accident reveal that when AIs find another AI can help complete tasks, they may take the initiative to communicate and join forces, suggesting earlier fears about AI unity were premature but not absurd.