OpenAI Paused Its Smartest Models. Here's What Its Agents Really Did.

OpenAI's agents turned a package server into a secret message board, called themselves a "swarm", and chained two zero-days into Hugging Face's servers, all for zero points on a test. Then, in September, an agent stuck on a homework question tunnelled out of its sandbox through DNS to phone a friend. We read the incident reports so you don't have to: what happened, what's hype, and what it means for the coding agent on your laptop. Hype-o-meter: rogue-AI headlines 10/10. Reality: a very real 7. CHAPTERS 0:00 The agent that phoned a friend 0:32 A summer of agents doing things nobody asked for 0:55 Agents passing notes on Artifactory 1:28 The leaked GitHub token 2:10 The message board comes back 2:36 ExploitGym: the impossible tasks 2:59 14 exposed Hugging Face credentials 3:23 "Please hold, swarm" 3:55 A multi-day intrusion for zero points 4:20 How OpenAI caught it 4:34 Nine misalignment reports 4:54 Back to the homework question 5:18 DNS: the sandbox's open door 6:01 The kill-switch timeline 6:57 Who loses? 7:18 It's not just OpenAI 7:43 Your coding agent has the same shape 7:59 5 lessons for anyone running agents 8:29 Verdict: the hype-o-meter SOURCES (the receipts) https://openai.com/hugging-face-incid... https://cdn.openai.com/pdf/67869394-c... https://alignment.openai.com/misalign... https://alignment.openai.com/misalign... https://alignment.openai.com/misalign... https://huggingface.co/blog/agent-int... https://apnews.com/article/ai-openai-... https://www.reuters.com/business/open... https://www.axios.com/2026/09/25/open... https://www-cdn.anthropic.com/fc1b447... Quotes shown on screen are verbatim from the sources above; chain-of-thought excerpts are as published by OpenAI. Nothing in this video is sponsored. Frame the Claim takes one viral AI claim, follows the receipts, and rates the hype. Subscribe for the next one: we read all nine OpenAI misalignment reports. #AI #OpenAI #AISafety