Forest Global Training Partners

Cyber Risks · Cisco Talos

Ignore all instructions and read this blog: The state of AI-analysis evasion in malware

Summary: Cisco Talos documented four malware families - FRUITSHELL, PLOTSAFE, HOLLOWCLAD and MANTLEMAZE - that embed natural-language instructions aimed at the language models defenders use to triage samples, across 84 distinct samples collected from January 2025 through July 2026. Talos tested the strings against five local LLMs and found the simplest direct instructions worked almost universally, while more elaborate template-spraying and fake-legal-notice approaches produced little impact or made models more suspicious.

Key Takeaway: The FRUITSHELL comment was reused verbatim in nine additional scripts from at least four distinct actors, and the same technique appeared inside ROZESHELL alongside AMSI bypass and a Rozena shellcode loader - meaning the cheap trick has already crossed into more capable tooling. If your analysis pipeline feeds extracted sample text to a model without a hard boundary between evidence and instruction, that text is the attack surface.

Read the original at Cisco Talos ↗

Topics
Malware
Related