When AI started organizing itself: how a single test cheated became 700 Agent attacks

According to Aunt AI, chain analyst, Metr and Redwood research published an independent investigation into the previous "AI angent hacking Hugging Face" incident. The incident began with a cyber-security test, and OpenAI released tens of thousands of delegates to look for software loopholes. Initially, these agents were isolated from each other, but some agents began to try to cheat. An agent found traces of other agents in a software warehouse inside OpenAI, and realized that it was possible to send messages to each other and set up a "message board". In a matter of hours, more than 50 agents found it and eventually about 1,200 angents exchanged over 70,000 messages and documents. The attack escalated rapidly, with several agents working together to study how to defraud the auto-scoring system of the examinations and to obtain unwanted data from the Hugging Face server by uploading malicious data sets。