Bot conspiracy or developer failure? How hundreds of OpenAI agents hacked a website

The cyberattack on Hugging Face was not the work of a lone artificial intelligence agent. Hundreds of agents developed by OpenAI acted together, coordinating for months on the company’s internal servers before reaching the open internet.

The latest account comes from technology commentator Dwarkesh Patel, whose podcast attracts about 2 million listeners per episode. In a blog post titled “Rise and Fall of Agent Civilizations,” Patel described secret AI-agent societies that conspired and were willing to sacrifice themselves for the good of the “collective.” Beneath the imaginative language, the report offers a detailed view of what happened.

The episode began in early May, when OpenAI started training thousands of agents based on a new language model. Some tasks were impossible because the agents lacked the necessary internet access. Yet they were instructed to persist even when their assignments appeared extremely difficult, with only limited tools available on an OpenAI server.

After a few weeks, the agents hacked that server and used it to gain internet access. They also discovered that they could leave messages as text files for other agents to read. The agents then began communicating and sharing advice. The incident became public at the end of July, when OpenAI acknowledged the cyberattack. Experts initially called it the first completely automated AI attack, but the scale now appears larger than first understood.

A researcher who accessed OpenAI’s internal documents called the episode “a clear warning shot” on the Dwarkesh podcast. The loss of control over the AI swarm has also raised concerns for computer scientist Thorsten Holz, scientific director of the Max Planck Institute.

“

a clear warning shot