# Closing the Binary Blind Spot: Domain-Knowledge Injection for LLM Malware Triage in npm Packages

DevFeed: [Closing the Binary Blind Spot: Domain-Knowledge Injection for LLM Malware Triage in npm Packages](<https://devfeed.tech/articles/closing-the-binary-blind-spot-domain-knowledge-injection-for-llm-malware-triage-in-npm-packages-88720.md>)

Original publisher: [Read original article](<https://www.databricks.com/blog/closing-binary-blind-spot-domain-knowledge-injection-llm-malware-triage-npm-packages>)

Author: Databricks Threat Research

Published: 2026-10-16T15:37:00Z

Content type: article

Language: en

Sources: [Databricks](<https://devfeed.tech/sources/databricks.md>)

Topics: [npm security](<https://devfeed.tech/topics/npm-security.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [AI software supply chain security](<https://devfeed.tech/topics/ai-software-supply-chain-security.md>)

Tags: [adversarial](<https://devfeed.tech/tags/adversarial.md>), [attacks](<https://devfeed.tech/tags/attacks.md>), [cybersecurity](<https://devfeed.tech/tags/cybersecurity.md>), [llm](<https://devfeed.tech/tags/llm.md>), [malware](<https://devfeed.tech/tags/malware.md>), [npm](<https://devfeed.tech/tags/npm.md>), [research](<https://devfeed.tech/tags/research.md>), [security](<https://devfeed.tech/tags/security.md>), [security-and-trust](<https://devfeed.tech/tags/security-and-trust.md>), [security-threat-research](<https://devfeed.tech/tags/security-threat-research.md>)

## AI overview

This article evaluates whether large language models can identify malware hidden in compiled binaries bundled with npm packages. In tests across nine models, injecting precomputed capability profiles encoding binary attack knowledge recovered 32 of 35 budget-tier detection failures and reduced failures at the escalation threshold from 50% to 4%. The authors also report a 4.25% false-positive rate on their benign validation set, while noting that production metrics use package-level OSV labels and hosted model results may drift.

## Source excerpt

IntroductionSupply chain attacks on package registries have grown in volume and sophistication [1]...