According to findings shared by ESET, the attacker’s strategy involves deliberately injecting content intended to trip an LLM’s safety mechanisms. The goal is to cause the AI tool to halt processing or refuse further interaction, effectively blinding the analyst to the malware’s true purpose.

• The technique relies on embedding prompts related to nuclear weapons or other catastrophic scenarios within the malware’s code or command output.
• When an AI-assisted analysis tool processes the file, the embedded prompt triggers a refusal or shutdown response, preventing full examination of the threat.

This tactic highlights a growing shift in adversarial behavior: instead of evading AI detection, some groups are now attempting to weaponize the constraints of those same systems. By forcing the AI to err on the side of caution, UAC-0099 can stall incident response efforts and buy time for their operations to proceed undetected.

It is worth noting that this is not a vulnerability in the AI model itself, but rather an abuse of its built-in safety protocols. The attack does not require exploit code; it simply relies on the predictable behavior of a well-trained model when confronted with highly sensitive or dangerous topics.

The campaign appears narrowly focused for now, but the technique could easily be repurposed by other groups looking to blind automated defense tools.

Source: The Hacker News

Has your team encountered any cases where malicious files deliberately caused your AI-powered security tools to stop mid-analysis, and how did you work around it?