CASHEWS: Source Preprocessor for LLM-based Malicious Package Detection
CASHEWS preprocessor boosts LLM-based malicious npm package detection, raising coverage to 98.8-100% and cutting false negatives by up to 18.6 points.
Researchers present CASHEWS, a JavaScript preprocessor for LLM-based malicious package detection that deobfuscates code iteratively, extracts bundled modules and dynamically executed code, identifies malicious sinks, and computes backward slices to produce compact detector input. Threat actors evade LLM detectors by exploiting limited context windows with high token-density obfuscation and by bundling malicious code with benign packages, as seen in supply-chain attacks such as Shai-Hulud. Across 512 large package files, two scanner types, and three LLMs, CASHEWS raised analysis coverage from 69.1-85.7% to 98.8-100% and reduced false-negative rates by up to 18.6 percentage points. Median preprocessing time is 30 seconds while net analysis cost drops 34.6%.
- Iterative deobfuscation, bundle extraction, dynamic code extraction, sink detection, and backward slicing compact JavaScript before LLM analysis.
- Coverage rises from 69.1-85.7% to 98.8-100% across 512 files, two scanners, three LLMs.
- Median preprocessing takes 30 seconds while cutting net analysis cost by 34.6%.
- Motivated by npm supply-chain attacks such as Shai-Hulud exploiting LLM context limits.
From the original article
1 min read at arxiv.orgMalicious npm package detection tools now leverage LLMs' semantic understanding of source code to detect malicious intent at scale. This capability has proven invaluable in identifying packages involved in recent supply-chain attacks such as Shai-Hulud.
By Jean-Charles Noirot Ferrand · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.