A single git push can leak billions of dollars in trained model value. PreserveIP detects model weight files in commits, forks, and pull requests — and alerts your security team before the window closes.
Model weight files committed directly to a git repository — the most common leak vector. Detected by file extension, size signature, and hash match.
Employee forks an internal repo to a personal account, then pushes weight files. PreserveIP tracks fork events and monitors the downstream repo.
.zip, .tar.gz, and .7z archives containing model files are flagged. Archive contents are inspected by filename and size before decompression.
Slack messages or PR comments containing signed S3 URLs, Google Drive links, or HuggingFace Hub references to proprietary models.
Public HuggingFace repository monitoring checks for model cards and weight files that match your registered fingerprints.
Fine-tuned derivatives inherit structural signatures from base models. PreserveIP's structural fingerprinting detects fine-tunes even when weights differ.
90-day pilot program available for design partners. No commitment required.