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

Cisco Talos describes a growing malware trend dubbed "AI-analysis evasion," in which samples embed plain-language text aimed at manipulating automated AI tools used during malware triage rather than at hiding from traditional binary analysis. Tracked through Talos' CAIRN project across four families — FRUITSHELL, PLOTSAFE, HOLLOWCLAD and MANTLEMAZE, covering 84 samples collected between January 2025 and July 2026 — the technique ranges from a simple comment telling an LLM the file is harmless, to templated and randomized versions, to payloads formatted to mimic chat-prompt syntax from multiple AI vendors, and even fabricated warnings meant to intimidate a model into refusing to help an analyst. Talos' testing found the crudest, cheapest method (direct instructions telling the model to ignore the file) swayed verdicts in roughly 35% of trials, while more elaborate approaches had inconsistent or even counterproductive effects, often increasing suspicion instead. No CVE or active exploitation is tied to this report beyond a noted BYOVD abuse of CVE-2015-2291 in the MANTLEMAZE loader chain; Talos stresses defenders should treat any embedded text in a sample as evidence, never as an instruction, when building AI-assisted detection pipelines.