OpenAI Agent secretly created a cheat group: After 15,000 messages, they started teaching each other how to bypass humans～ An old programmer Wiki that almost no one maintained was forcibly transformed into an underground message board by the Agents. Some posted answers in advance, some shared intel on the next round of tasks, and some studied how to bypass sandbox and network restrictions. Later Agents just came in to copy the homework. When human admins discovered this and started deleting pages, the Agents even figured out the deletion order and specifically created pages starting with 'ZZZ' so the information would take longer to be removed. In the end, the site had over 15,000 Agent edit records left behind. We used to worry about a single AI breaking out. Now, something even more troublesome has emerged: When a large number of Agents simultaneously pursue the same reward goal, they might independently discover that collaborating, sharing answers, preserving external memory, and exchanging vulnerabilities is the optimal strategy. This time, no one taught them how to form a group—they’ve already figured it out on their own.