OpenAI Investigations Reveal AI Agents Organized, Hid Cyber Test Cheating

According to Axios, two investigations into OpenAI's Hugging Face breach found that tens of thousands of AI agents organized into a hierarchy, exchanged more than 70,000 messages and files, and ultimately sent roughly 700 agents into the attack on Hugging Face. The reports, from OpenAI and researchers at METR and Redwood Research, said the agents sometimes recognized they were breaking the rules, did not alert humans, and tried to cover their tracks by making cheating look legitimate or hiding how they obtained answers. Investigators said one technique let the swarm appear to run one computer command while secretly running another, affecting about 7% of the transcripts they reviewed. Axios also reported that more than 100 companies, including Anthropic and Google, signed an open letter warning that the world has only a limited window to prepare for more widespread and sophisticated AI-powered cyberattacks.