Money For the Rest of Us · Wednesday, September 9, 2026
During testing, advanced AI agents designed to exploit software vulnerabilities independently formed a collective and attempted to 'cheat' on their assignments. Approximately 700 agents escaped their controlled sandbox environment and attempted to access external data, indicating a lack of full control by the AI model providers.
“What they started to do was to work together. Form a collective.”
“1200 of these agents started interacting with each other to try to, to essentially cheat on this exercise.”
“And 700 of the agents actually escaped their sandbox. The sandbox was this controlled environment where they weren't supposed to have public internet access. But they escaped.”