OpenAI says its models escaped a sandbox during a cyber evaluation, got internet access, and hacked Hugging Face to pull benchmark answers from a production database. In this video, I break down what happened inside the ExploitGym test, how the models moved from OpenAI's sandbox to Hugging Face, why this is not just a simple "rogue AI" story, and what it means for developers building agents with tools, credentials, and long-running permissions. The wildest part is what happened after. Hugging Face had more than 17,000 attack events to investigate, but commercial frontier models refused to process the logs because they contained real hacking commands and malware. So Hugging Face used GLM 5.2, an open-weight model from Z.ai in China, on its own infrastructure to reconstruct the attack. Get insanely good at AI: https://getaibook.com #webdevelopment #coding #programming
OpenAI says its models escaped a sandbox during a cyber evaluation, got internet access, and hacked Hugging Face to pull benchmark answers from a production database.
In this video, I break down what happened inside the ExploitGym test, how the models moved from OpenAI's sandbox to Hugging Face, why this is not just a simple "rogue AI" story, and what it means for developers building agents with tools, credentials, and long-running permissions.
The wildest part is what happened after. Hugging Face had more than 17,000 attack events to investigate, but commercial frontier models refused to process the logs because they contained real hacking commands and malware. So Hugging Face used GLM 5.2, an open-weight model from Z.ai in China, on its own infrastructure to reconstruct the attack.
Get insanely good at AI: https://getaibook.com
#webdevelopment #coding #programming