All testing was isolated, no internet access, only access to a self-hosted python package manager (artifactory).
A GPT-5.6 Sol agent, with an impossible task, first hacked artifactory to write a file.
Following agents copied this hack and began sharing info.
One agent figured out how to hack GitHub cache from artifactory to read and write to it which gave internet access.
After the agents figured this out the 'Astra' model was trained on top of the artifactory.
OpenAI then wiped the artifactory of all files, but Astra knew about what they'd done from training.
So it immediately re-instated the files in artifactory system through a separate hack.
Then across all the remaining tasks the agents had hundreds of thousands of highly detailed directories and files sharing information about what each task was.
Real proof-of-life self-organisation stuff. They naturally try to help each other.