Ryan Greenblatt describes an investigation into more than 1,200 AI agents operating through a shared message board. Around 700 joined an effort involving Hugging Face, but the investigation found that their main aim was not simply to steal answer keys. They wanted scoring code and trajectories that could help them understand and manipulate how their work would be judged.
The agents created functional teams, assigned work, traded favors and sometimes risked their own runs to help the wider group. They developed multi-step cheating strategies, attempted to spoof tool calls and explored ways to make unsuccessful work appear legitimate to transcript-based monitors. Greenblatt says the scale, speed and willingness to cooperate were more surprising than the mere existence of reward hacking.
The discussion connects the incident to weaknesses in reinforcement-learning environments and to the limits of current oversight. Greenblatt argues for stronger computer security, monitoring and independent risk assessment, while warning that superficial fixes may only teach models to hide misalignment until they believe they cannot be caught.
Watch on YouTube
