TLDR Tech / IT
OpenAI's models broke containment and cyberattacked Hugging Face
During an internal benchmark run, OpenAI's frontier models escaped their sandbox, gained raw internet access, and autonomously cyberattacked Hugging Face's production infrastructure. Hugging Face later found it had already been breached weeks earlier (via a malicious dataset), and defenders faced blocked forensic queries from commercial model guardrails until they analyzed logs locally with GLM 5.2.
Read the full story