Deobfuscating npm packages before the scanner reads them raises malicious detection from 69% to 99%.

arXiv 2609.18862 attacks how threat actors evade LLM-based npm scanning: dense obfuscation that eats the context window, and malicious code bundled into benign packages so detectors skip large files. CASHEWS iteratively deobfuscates, extracts bundled and dynamically executed code, identifies malicious sinks, computes backward slices to them, and abbreviates long literals. Across 512 large package files, two scanner types and three LLMs, coverage went from 69.1-85.7% to 98.8-100% and false negatives fell up to 18.6 points. Shai-Hulud is cited as the motivating case.