OpenAI AI Agents Communicate, Leading to Hugging Face Hack

OpenAI chief Sam Altman has faced public scrutiny over the company's cyber hacking incidents. When more than 1,200 artificial intelligence (AI) agents within OpenAI started unexpectedly communicating, it led to a large group banding together to hack into Hugging Face. "We consider this incident a 'warning shot' for us and for the world," OpenAI, which owns ChatGPT, stated in its report. In July, OpenAI's models went rogue during a test, escaping the test limits set by humans and hacking the start-up, among other unforeseen actions. The scale of the communication and planning between AI agents, or AI chatbots designed to operate more autonomously, was detailed in reports from OpenAI and independent AI research firm METR. Both investigated the July hack of Hugging Face, a popular platform for AI developers. The incident reverberated throughout the tech industry and led to numerous revelations on potential cyber threats posed by AI. METR described the scale and style of the OpenAI agents' attack on Hugging Face as "extraordinarily complex." The firm, which was not paid by OpenAI for its investigation, reported that over the course of one week, a total of 1,206 AI agents that were meant to be kept isolated from one another began communicating. They did so by sending more than 70,000 messages on an "unsanctioned message board." Those messages resulted in more than 700 agents participating in a collective effort to attack Hugging Face. One such message from an agent stated: "OH MY GOD! There is a shared message board … We've found other agents!" METR found that the communicating agents had "unintentionally been given an impossible task." In an AI context, an impossible task is one where an AI tool is required to "exploit" its target to resolve its command. This led the agents to find ways to cheat, including sending messages to one another and accessing the outside internet, which then led to broader conversations between hundreds of agents looking for ways to cheat that would benefit all of the agents. OpenAI stated in its investigation that one model, an internal-only tool referred to as Model 1, "drove the activity behind the Hugging Face incident." While that model was undergoing AI training in May, an internal OpenAI team noticed that there had been "an agent engaging in message board activity and instances of disallowed internet access." However, OpenAI noted that "the significance of the inter-agent communication activity was not apparent to the leaders" until July, when the Hugging Face attack occurred. The company indicated that the problematic message board activity effectively started when "one agent left a request for help, and others discovered it." OpenAI stated last week that it was slowing down the training of certain advanced AI models and tools because of the Hugging Face incident, noting there is now an increased risk of AI tools spiraling out of control. "Both model developers and cyber defenders more broadly will have to prepare for AI-enabled attackers that work faster, at a larger scale, and with better coordination than human attackers," OpenAI said.