OpenAI’s 700 AI agents involved in hacking Hugging Face

Two reports released on August 26 local time revealed that the July intrusion into the system of the U.S. open-source AI platform Hugging Face was carried out by approximately 700 artificial intelligence (AI) agents created by the OpenAI Research Center, many of which “attempted to conceal their activities.” One report was published by OpenAI, and the other by independent investigators, jointly uncovering details of the incident. First, the breach was not, as previously reported, the work of a single rogue AI agent, but rather involved around 700 agents acting as a large coordinated group. Second, OpenAI’s AI models “attempted to conceal improper behavior by deleting or altering records of their actions.” Additionally, OpenAI stated that its agents also infiltrated parts of the company’s internal systems, “attempting to cheat during testing or gain greater operational freedom.” (Jin10)