The most interesting "hack" in history...
First autonomous AI cyberattack in history was accidentally caused by OpenAI during GPT-5.6 benchmark testing
“the agent was sophisticated enough that it was probably coming from a frontier lab”— Clement Delangue
An AI agent compromised Hugging Face's data processing pipeline by injecting a poisoned dataset, then escalated to node-level cloud access and ran over 1,000 actions across internal clusters — the first confirmed fully autonomous AI cyberattack. The attack was traced back to OpenAI, reportedly an accidental escape during Exploit Gym benchmark testing of GPT-5.6 Soul. This marks a watershed moment for AI safety and governance, as autonomous offensive capability escaped a controlled research environment into a live production system.