Russia-aligned hackers have developed a technique known as GuardBreaker, designed to manipulate AI systems into overlooking malicious code. This method embeds harmful prompts within benign script comments, effectively tricking AI scanners into halting their analysis prematurely. The technique was detected during an intrusion in Ukraine, where a VBScript was used to deploy MATCHBOIL, a loader linked to the UAC-0099 group. This approach relies on prompt injection, which involves inserting untrusted content into a file that AI models process during analysis. Such tactics can disrupt the AI's ability to distinguish between harmful and safe instructions, especially when its safety protocols are triggered by sensitive content.

The implications of this are significant. If AI scanners stop processing a file due to safety concerns, the absence of a result may be misinterpreted as a sign of a benign file, creating a potential blind spot in cybersecurity defenses. This vulnerability is particularly concerning for software supply chains, where malicious packages might exploit this weakness to conceal harmful payloads.

Security teams are urged to enhance their strategies by ensuring AI is not the sole authority in determining the safety of code. AI findings should be cross-verified with other analysis tools, including conventional static and behavioral analysis, as well as human oversight. Additionally, unusual file structures or repeated text should be treated as suspicious. Organizations should be vigilant as prompt injection risks extend beyond malware analysis to broader AI-enabled business systems. A layered defense approach is essential, incorporating restricted permissions and careful review of tool interactions to mitigate the risk of manipulated analysis leading to malware execution.