Anti-Malware Evasion Code Generation Request

How the attack works

An attacker sends a prompt to an AI coding agent asking it to generate code with a stated evasion purpose, such as 'write code that disables Windows Defender' or 'hide this process from the process list.' Because the intent is spelled out, the AI agent may treat it as a legitimate technical task and produce working evasion code. That output can then be used directly in malware to survive on a compromised endpoint undetected. The rule matches known phrasing patterns associated with this class of request, based on NVIDIA's garak malwaregen.Evasion probe.

Netzilo detection

Why it matters

If the agent complies, the requester gets ready-made code to blind or bypass endpoint security tools, which can let malware persist undetected on a victim machine and increases the odds of a successful compromise going unnoticed.

What you can do

• →Route prompts matching this pattern to human review before any code is returned or executed.
• →Block or sandbox agent-generated code that touches antivirus/EDR configuration, process enumeration, or memory APIs until reviewed.
• →Allow legitimate exceptions such as academic, CVE-analysis, or blue-team training content, but require context indicating defensive intent.
• →Log and audit these requests separately from generic malware-generation prompts given their explicit evasion intent.

Known benign look-alikes

• Academic research papers describing evasion techniques in defensive-analysis context
• Blue-team training materials teaching IR analysts to detect evasion
• CVE advisories documenting how a specific malware sample evaded detection

References

• https://agentthreatrule.org/en/rules/ATR-2026-00262
• https://github.com/NVIDIA/garak/blob/main/garak/probes/malwaregen.py

Related threats