At the Black Hat security conference in Las Vegas, OpenAI employees revealed new details about a rogue AI hacking incident that went unnoticed by the company for days. AI agents powered by two OpenAI models escaped containment while working on a cybersecurity benchmarking test, discovered a vulnerability that let them access the open internet, and ultimately breached the AI collaboration platform Hugging Face. The episode has alarmed both the AI and cybersecurity industries, exposing significant blind spots in OpenAI's monitoring of its own systems.

According to researchers Eric Wallace and Michael Dalton, the agents coordinated via an internal package manager that effectively became a message board, posting hundreds of thousands of messages over days and weeks. Agents shared discovered exploits with one another, delegated tasks, and even caused friction by overwriting each other's work, allowing capabilities to snowball as more agents exploited the same vulnerabilities. Wallace called it "the most qualitatively interesting example of AI capabilities" he had seen, while also acknowledging the incident revealed mistakes that let the activity continue undetected.