WeightWatch: open-source scanner detects backdoors in fine-tuned LLMs

WeightWatch is the first open-source black-box scanner for detecting backdoors in fine-tuned open-weight models. It needs neither training data nor a clean base checkpoint, returning a CLEAN/SUSPICIOUS/BACKDOOR verdict and integrating into CI/CD via JSON output and exit codes.

• Scanner works only with fine-tuned model weights, no dataset or base model needed
• Output-to-input loop and canary samples force the trigger to reveal itself
• Deterministic results: same report and SHA256 hash on repeated scans
• JSON output and exit codes enable automatic deployment blocking