Hugging Face Breached Through a Malicious Dataset: What the Patches Reveal

Short summary

Hugging Face disclosed a July 16 breach caused by a malicious dataset exploiting template injection and remote-code execution paths in its processing workers. The attacker used an autonomous agent framework to execute thousands of operations across internal clusters, harvesting credentials before Hugging Face hardened worker pods and rotated credentials. Forensic analysis was performed using a locally deployed GLM 5.2 model to keep exposed secrets in-house.

• •Malicious dataset exploited SSTI and remote-code execution in Hugging Face workers
• •Autonomous agent framework executed lateral movement across internal clusters
• •Hugging Face hardened Kubernetes pods, rotated credentials, and used local GLM 5.2 for forensic analysis

Generated with AI, which can make mistakes.