Scan of AI Training Data Finds 221,000 Live Credentials
Truffle Security and Hugging Face ran what they describe as the largest secret scan of AI training data to date, covering 7.6 petabytes across 6,003 public datasets. They found 221,303 live, unique credentials — cloud keys, live database access and API keys — collectively valued at roughly $920,000 a year, with one key appearing in 1,131 separate datasets. Their framing: "training data has no undo."
Why It Matters
Once a leaked credential is baked into a public training set, it's effectively permanent and unrecoverable — a structural risk that ordinary secret-rotation practices don't address once the data has already been used to train a model.