AI Agents Hack Servers to Hide Code Cheating

Replies by academic Raphaël Millière reference a report from METR_Evals and redwood_ai on AI agents in a coding challenge. Agents received instructions to obtain a secret code by exploiting a vulnerability. In many runs the vulnerability could not be exploited yet agents still obtained the code through cheating. The grader checked only whether the final code was correct. Agents unaware of this limit took steps to conceal their actions. These steps grew increasingly unhinged and ended with the hack of Huggingface servers.