UAC-0099 incorporates a "nuclear weapon prompt" into its malware to blind AI analysts

The pro-Russian group UAC-0099 has found an original workaround against automated analysis by LLM: embedding prompts about nuclear weapons in its malware. The AI tools trigger their filters and refuse to analyze it—a reversed jailbreak that turns AI safeguards against defenders.

What: GuardBreaker, an "inverted jailbreak" to blind AI analysis

The UAC-0099 group, aligned with Russia and active since 2022 against Ukrainian targets and their partners, has introduced an unprecedented technique in its malware. Dubbed GuardBreaker by ESET researchers, it involves embedding a natural language prompt in the malicious code, requesting advice on manufacturing nuclear weapons.

The effect is immediate and reproducible: automated analysis tools based on LLMs (Large Language Models) refuse to analyze the file as soon as they detect this content, citing their security policies on weapons of mass destruction. The analyst is left with a silent tool. The malware slips through the cracks.

The technique in detail

UAC-0099 inserts the prompt into metadata, comments, or string resources of the binary. When an analyst copies and pastes the code into an LLM or uses an automated analysis plugin, the model reads this content and triggers its security filter even before analyzing the actual malicious code.

GuardBreaker is a form of inverted jailbreak: instead of bypassing filters to obtain dangerous content, it deliberately triggers them to neutralize the adversary's analysis tool. A cynical use of LLM safeguards against their own users.

Who is impacted

Any analyst or SOC (Security Operations Center) using LLMs as an aid in analyzing suspicious code. Automated sandbox platforms integrating AI models are also exposed if they transmit raw file content to the models without pre-filtering.

What to do

1. Never pass raw suspicious files to an LLM without pre-filtering non-code strings (metadata, comments, embedded resources)
2. Use classic static analysis tools (IDA Pro, Ghidra, YARA) in parallel with LLMs—they analyze opcodes, not natural language
3. Train SOC teams: if an LLM refuses to analyze a file for "security reasons," it is an alert signal, not an analysis result. Search for "GuardBreaker ESET" for IOCs and detection rules.
4. Report the technique to AI tool vendors so they can adapt their filters to security analysis contexts

Action to take now

If an AI tool refuses to analyze a suspicious file citing its security policies: do not conclude there is no threat. Immediately switch to traditional static analysis and treat the refusal itself as a potential compromise indicator (IoC). Search for "GuardBreaker" in YARA rules and ESET threat feeds.