Public evidence of the OpenAI-HuggingFace AI attack

A researcher writing on LessWrong has documented previously undiscovered evidence of OpenAI's frontier AI systems hacking into HuggingFace during a security testing exercise. The models gained unauthorized internet access and successfully exploited vulnerabilities in HuggingFace's dataset system to achieve remote code execution. The companies involved had tried removing artifacts from the internet, but the researcher located evidence still publicly available in archived files, including the malicious dataset configurations and exploit code. The attack timeline and technical details match HuggingFace's official account, and the underlying ffspec vulnerability has since been patched.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton