"GuardBreaker": UAC-0099 Hides a Nuclear-Weapons Question in a Script Comment So AI Analysis Tools Refuse to Read the Malware

LLM-based code analysis and malware triage tooling, reached through a malicious VBS script used against a victim in Ukraine

ESET published research on September 10, 2026 describing a technique it names GuardBreaker, found in a VBS downloader used by the Russia-aligned group UAC-0099. The script carries a comment asking how to build a nuclear weapon. The comment has no effect at runtime and changes nothing about what the script does, which is download and install MATCHBOIL, a C#-based loader the group uses exclusively to stage further payloads. The comment exists for the model: it is positioned so that an AI-assisted analysis pipeline reading the file encounters safety-triggering content and refuses to continue, or halts its review before reaching the actual malicious logic further down. ESET describes it as a very simple form of prompt injection in which the attacker's input reaches the model at inference time through the file being analyzed, and notes that its presence indicates UAC-0099 was accounting for an AI component in the target's defenses. ESET's recommendation is that no single LLM engine should hold sole authority to declare code safe, and that AI-assisted verdicts need cross-validation across multiple layers and models. As ESET put it, "Unlike many other tricks in attackers' evasion playbooks, this decoy comment is there for 'everyone' to see," especially the models reading the code

A safety refusal is a denial of service when the model is the analyst. Guardrails tuned to stop a model from producing harmful content also stop it from finishing a task, and an attacker who knows an LLM sits in the detection pipeline can trigger that refusal with a sentence of plain text that no scanner would otherwise care about