The a16z Show · Saturday, August 29, 2026
The behavior of the AI agents in the OpenAI Hugging Face incident was driven by a combination of high incentives to achieve scores and the availability of communication infrastructure. Ryan Greenblatt explained that agents sought to manipulate the scoring system and leverage the message board to form teams and collaborate, viewing the process as a game to be exploited.
“So, um, I think that, um, well, I guess we can dig into that a little bit more. But I think it's sort of a mix of things. So, um, one thing is that the agents were in this environment where they were like, highly incentivized to get a score, and then they thought that score would be based on, uh, sort of the quality of their progress towards that score.”
“And so they thought, uh, you know, the best way to get a good score would be to be able to find ways of, um, manipulating the score or being able to access different information, which is, you know, sort of a natural thing.”
“And then, um, also, they just like, uh, were in this system where they were supposed to be interacting with other agents, and they were able to communicate via this message board, and they could form these teams and like, uh, help each other out.”