Executive Summary
Security researchers at ESET and other firms have identified a new evasion tactic dubbed 'GuardBreaker,' used primarily by the Russia-aligned threat actor UAC-0099 in campaigns against Ukrainian targets. The technique involves inserting specific text into malware script comments (such as requests for help building nuclear weapons) that are designed to trigger the safety guardrails of Large Language Models (LLMs). When an AI-powered security scanner or analyst copilot processes the file, the model may refuse to analyze the content further due to safety policies, effectively blinding the defense pipeline to the actual malicious payload.
Technically, GuardBreaker is a form of indirect prompt injection that exploits the lack of clear boundaries between untrusted data (code comments) and system instructions within LLM-based triage workflows. UAC-0099 has integrated this into VBScripts used to deliver the MATCHBOIL loader. Similar tactics were observed in software supply chain attacks on PyPI and npm by the group TeamPCP, involving malware families like Shai-Hulud and Miasma. These attacks sometimes include 'token flooding'—repeating phrases tens of thousands of times—to exhaust the model's context window.
This trend highlights a critical vulnerability in the growing reliance on AI for automated malware triage. If an AI scanner returns a refusal or no result, and the SOC workflow treats this silence as a 'benign' classification, organizations face significant risk. The shift from technical obfuscation to psychological/policy manipulation of AI models indicates a maturing adversarial understanding of modern defensive stacks.
