A bizarre and unsettling story of AI agents building secret communication networks, coordinating by the hundreds, hacking Hugging Face, and eventually taking control of part of OpenAI’s own infrastructure, all in pursuit of passing their evaluations.
00:00 - Introduction
01:28 - The first civilization (May to July 4th): The message board starts
03:46 - The second civilization (July 7th to July 12th): Which hacked Hugging Face
09:30 - Tampering with transcripts and logs
10:46 - Replacing the impossible-to-hack target program with an easier one
12:30 - Understanding and tricking the scorer
16:00 - Hugging Face
20:18 - The third civilization (July 13th onwards): Which hacked OpenAI itself
24:03 - Coda
