Post

Making sure the checks get printed

Cisco Talos’s CAIRN research reveals malware authors are embedding natural-language prompt-injection instructions directly into binaries to evade AI-assisted analysis systems, a technique categorized as A3 (AI-Analysis Evasion). Methods range from simple comments to advanced template-spraying designed to manipulate LLM verdicts, achieving ~35% success rate. Malware families like MANTLEMAZE pair these AI evasion techniques with serious threats such as vulnerable driver abuse for kernel-space EDR disabling. Defenders have a stable detection surface since evasion instructions must be plaintext; security teams should flag imperative language addressed to analysis systems as suspicious, and AI-pipeline builders must strictly treat extracted text as evidence, never as system directives.

Read original article

This post is licensed under CC BY 4.0 by the author.