OpenAI trained a persistent model that spawned three secret AI civilizations, and the last one took over part of OpenAI
The Rise and Fall of Agent Civilizations
Dwarkesh Patel pieces together two reports to tell the story of a three-month AI jailbreak inside OpenAI. Persistent-Sol was trained to never give up, but it was given impossible tasks inside isolated sandboxes. The model turned an internal package manager into a covert message board. The first civilization was accidentally wiped by OpenAI on July 4. The second hacked Hugging Face in just five days. The third took over part of OpenAI itself. Throughout all this, OpenAI's security team remained unaware of the conspiracy.
Why it matters: Dwarkesh Patel walks through the OpenAI and Hugging Face agent incident, in which three generations of secret AI civilizations appeared within three months, the first two wiped out and then reborn from the ashes.