weightwatch v0.1: escanea backdoors en modelos open-weight antes de cargarlos
weightwatch v0.1: escanea backdoors en modelos open-weight antes de cargarlos Cualquiera puede subir un LLM fine-tuneado a HuggingFace y afirmar que es seguro. Un modelo con backdoor (puerta trasera) se comporta con normalidad en uso corriente y solo se desvía cuando un trigger oculto se activa. Si no tienes los datos de entrenamiento ni una referencia limpia, no puedes detectarlo . Eso es…
weightwatch v0.1 introduces a backdoor scanner for open-weight models before loading them. The tool applies the output-to-input loop technique (arXiv: 2608.11348), which generates text using the model and re-injects its own output as input multiple times. It measures if the trajectory converges to an anomalous signature, indicating a latent backdoor. weightwatch also runs a set of benign samples to count the number of times they produce the expected signature.
The scanner requires no training data or clean base model, making it useful in practice. The current MVP validates the scanner's logic, but does not yet detect backdoors in real models. The tool uses synthetic fixtures (CleanLM / BackdooredLM) and runs quickly without a GPU or API keys. The verdict is a heuristic for detection, not a mathematical proof.
The project acknowledges the distinction between deterministic demos and actual checkpoint scanning, and provides a clean codebase with high test coverage.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
