weightwatch v0.1: Open-Source Tool Scans Third-Party AI Models for Hidden Backdoors

Developer Pedro Sordo Martínez has released weightwatch v0.1, a black-box scanner designed to detect backdoors in open-weight AI models before they are loaded or trusted. The tool addresses a gap identified through research: while 75 arXiv papers from 2026 document the backdoor threat in fine-tuned models, virtually no open-source tooling exists to counter it. weightwatch works by repeatedly re-injecting a model's own output as input and monitoring whether the response trajectory converges to an anomalous signature, also firing a set of canary inputs typical of known backdoor triggers. It returns one of three verdicts — CLEAN, SUSPICIOUS, or BACKDOOR — without requiring access to training data or a clean reference model. The current v0.1 release validates the detection logic using synthetic fixtures rather than real HuggingFace checkpoints, with real-model scanning planned for v0.2.