Russian Hackers Inserted a Nuclear Weapon Request Into Malware to Fool AI

ESET described a new attack technique in which hackers don't try to break AI's safeguards, but rather the opposite - they deliberately trigger them. The Russia-linked group UAC-0099 pasted a sentence about wanting to build a nuclear weapon into malicious scripts, hoping that the AI system analyzing the code would halt its work before detecting the actual threat.

The mechanism is simple yet clever. AI systems used to analyze malware have built-in safety filters that react to content deemed sensitive, such as instructions related to weapons of mass destruction. When such a text fragment appears in an analyzed file, the model may refuse to continue working or shift its focus to neutralizing the supposed threat instead of the actual code.

A Reversed Jailbreak

ESET compares GuardBreaker to a reversed jailbreak. A classic jailbreak tries to get an AI model to ignore its safety rules and do something forbidden. GuardBreaker does the opposite - it deliberately activates those rules to paralyze the analysis and smuggle malicious code past detection systems.

This is meant to draw the AI's attention to security-sensitive content and stop it from analyzing the rest of the code - ESET

The researchers stress that the text pasted in as a comment serves no function in the script itself - its only purpose is to mislead the large language model-based analysis tool. Without this trick, the code itself could have been detected like any other malware.

Not the First Case

ESET notes that GuardBreaker is an evolution of previously observed tactics. Back in June 2026, researchers described attacks on the Python package supply chain in which attackers embedded similar content about biological and nuclear weapons in code to force AI-based scanners to refuse analysis. GuardBreaker is the first documented case of this method being used by a specific, known state-linked group - UAC-0099 - in a real campaign against critical infrastructure.

UAC-0099 has been active since at least 2022 and was previously associated mainly with phishing attacks on Ukrainian government institutions and companies in the transport and energy sectors. The use of the MATCHBOIL loader, previously described by CERT-UA as a tool distributed under the guise of fake plugins for the popular Notepad++ editor, fits the group's established pattern - infecting critical infrastructure employees through seemingly harmless tools.

AI's Limits in Security

Juraj Janosik, ESET's Vice President of Artificial Intelligence, warns against blindly relying on AI models for threat analysis.

AI and machine learning can be valuable tools in security, but they cannot be trusted unconditionally or treated as a panacea for every possible threat - Juraj Janosik, Vice President of AI, ESET

Janosik adds that if AI-assisted analysis isn't combined with a multi-layered approach that includes expert research, behavioral analysis, reputation systems, sandboxing, heuristics, telemetry, and human-driven engineering, attackers will look for ways to manipulate or bypass it.

For companies deploying security tools based on large language models, this is a warning sign. More and more antivirus products and SOC systems use AI for initial triage of suspicious files, which speeds up the work of security teams, but as the UAC-0099 case shows, it also creates a new attack surface - not on the infrastructure itself, but on the model's reasoning process.

The case also points to a broader trend known as AI evasion, techniques designed specifically to bypass automated AI-based systems. As more companies replace manual malware analysis with language models, this type of attack is likely to become more common, since it doesn't require breaking security in the traditional sense - just knowing how the model reacts to certain keywords.

ESET did not disclose which specific commercial tools or AI-based security products proved vulnerable to this technique, limiting its disclosure to describing the mechanism and attributing it to the UAC-0099 campaign in Ukraine.