Wiz, a cloud security company, has reported a serious vulnerability found on the Hugging Face AI model platform, allowing access to other users’ AI model data.
The security engineering team at Wiz utilized a method of reverse engineering by uploading their own malicious model onto Hugging Face and then exploiting the system at the container level, granting access to other users’ model data on the platform.
The two critical vulnerabilities discovered by Wiz that enabled this exploit were the AI Inference component commonly using the pickle format, allowing for the execution of unsafe commands, and the ability for CI/CD processes to run consecutively, making it possible to exploit the system and manipulate CI/CD clusters.
These vulnerabilities have been reported and addressed by Hugging Face, with Wiz announcing a collaboration to enhance security measures on the platform moving forward.
TLDR: Wiz identified and exploited vulnerabilities on the Hugging Face platform, leading to a collaboration to enhance security measures.
Leave a Comment