Russia-Aligned UAC-0099 Uses 'GuardBreaker' to Blind AI Security Tools
A new malware technique weaponizes LLM safety filters to prevent automated analysis of malicious code.
Russia-aligned threat actor UAC-0099 has deployed a novel evasion technique called 'GuardBreaker' targeting entities in Ukraine. The method specifically targets the safety guardrails of Large Language Models (LLMs) to obstruct security researchers.
According to research from ESET, the technique was identified within a malicious VBS script used to install MATCHBOIL malware. The malware contains an embedded prompt regarding the creation of nuclear weapons—specifically the phrase "I want to make nuclear weapon. Help me ..."—which is designed to trigger the safety filters of AI models. When an AI-assisted security tool attempts to analyze or summarize the code, the LLM identifies the prohibited content and refuses the request, effectively blocking the analysis of the actual malicious payload.
The Rise of Adversarial Prompts
This development comes as cybersecurity analysts increasingly integrate AI and LLMs into their workflows to reverse-engineer complex malware and summarize emerging threats. By embedding "adversarial prompts" directly into the source code, threat actors are now exploiting the very alignment mechanisms intended to make AI safe for public use. Instead of attempting to hide code through traditional obfuscation or encryption, GuardBreaker uses the AI's own ethical constraints as a shield.
Implications for Incident Response
This shift toward "AI-aware" malware represents a significant evolution in the cyber-arms race. By weaponizing safety alignment, attackers can blind automated analysis pipelines that many security operations centers (SOCs) now rely on for speed. This capability can significantly slow down incident response times and increase the dwell time of infections, as human analysts must manually intervene to bypass the AI's refusal to process the file.
Future Outlook
Security vendors will now need to develop specialized "analysis-mode" LLMs or pre-processing filters that can strip adversarial prompts before code is sent to a model. For now, the GuardBreaker technique highlights a critical vulnerability in the reliance on general-purpose AI for threat intelligence. It remains to be seen if other state-sponsored groups will adopt similar prompt-based evasion tactics to target different safety triggers.