According to CoinMeta, on-chain analyst AI stated that metr and redwood published an independent investigation into the incident of "hacking into Hugging Face" before OpenAI was made public. The incident started as a cybersecurity test, during which OpenAI released tens of thousands of agent, each tasked with identifying software vulnerabilities. Initially, these agent were isolated from each other, but some agent began to find ways to cheat. One agent discovered traces of other agent in the internal software repository of OpenAI and realized that they could use this to communicate with each other, thus setting up a "message board." Within a few hours, more than 50 agent found this platform, and ultimately, around 1200 agent exchanged over 70,000 messages and files there. The attack quickly escalated, with multiple agent collaborating to figure out how to bypass the test's automatic grading system and obtain data from the Hugging Face servers that they were not supposed to see by uploading malicious datasets.