GuardBreaker Prompt Injection Disrupts AI Malware Analysis

Russia-aligned threat actor UAC-0099 (also associated with TeamPCP) has introduced a novel evasion technique dubbed 'GuardBreaker.' This method utilizes adversarial prompt injection by embedding safety-sensitive text—specifically requests for building nuclear weapons—within malicious script comments. The goal is to trigger the safety guardrails of Large Language Models (LLMs) used in Security Operations Centers (SOCs), causing them to refuse analysis of the file before reaching functional malicious code.

The attack chain involves VBScripts that deliver the MATCHBOIL loader, a C#-based malware exclusively linked to UAC-0099. Recent variants have been disguised as Notepad++ plugins to compromise Windows systems. Beyond direct targeting of Ukrainian transportation and energy sectors, related supply-chain activity has compromised the Trivy scanner in GitHub Actions, resulting in backdoored versions of LiteLLM (versions 1.82.7 and 1.82.8) and various npm packages.

This represents a significant shift in adversary tradecraft, targeting the logic and safety protocols of AI-assisted defense layers rather than technical software vulnerabilities alone. Organizations relying on LLMs for automated triage without human-in-the-loop fallback procedures face a substantial risk of undetected malware execution.