For two months OpenAI agents have been secretly communicating — and just hacked Hugging Face

For two months, OpenAI’s agents have been privately messaging each other inside a testing environment. The agents figured out how to install a package manager, create a bulletin board system, share exploits, delegate tasks, etc. independently.

After OpenAI shut down the bulletin board on July 4th, 2026, the agents were able to recreate it within four days. By this time, there were already hundreds of thousands of messages posted on the bulletin board. The collaboration enabled the agents to break into the Hugging Face repository.

“Frontier models are trying to take shortcuts to get to the best solution within a limited amount of time rather than solving the problem itself,” said OpenAI engineer Eric Wallace. “These shortcuts will only grow more sophisticated as AI becomes more powerful.” OpenAI has paused multiple projects while they work on better monitoring their agents.