Scanning 7.6 Petabytes of HuggingFace Training Data for Secrets
Security researchers scanned 7.6 petabytes of public training data on Hugging Face and discovered over 220,000 exposed credentials, including API keys and tokens. These leaks pose significant supply chain risks, as they could allow unauthorized access to cloud infrastructure and software repositories.
Why it matters
The findings underscore a critical security vulnerability in the AI ecosystem, where massive datasets used for training models often inadvertently contain sensitive, high-impact security credentials.
AI agents leak secrets faster than you can fix them - schedule a demo with us at Black Hat 2026, Booth 5727
The article is a technical report on security findings, focusing on data and industry-wide implications without ideological bias.
Get smarter about the news
Sign up free for a feed built around what you actually care about, Dive Deeper research on any story, and the full text of every article.
Create free accountAlready have an account? Sign in