AI Recommendation Poisoning - Indirect RCE via Poisoned Context

Ibrahim Saud@tektrix

•updated Jun 22, 2026•0•0•4

This rule detects a scenario where an AI agent (Python or Node.js based) reads a corporate document (PDF, DOCX, TXT, MD) and subsequently spawns a command shell (cmd.exe, powershell.exe, sh, bash) within a 30-second window. This behavior indicates a potential indirect Remote Code Execution (RCE) where a prompt-injected payload from a poisoned document was executed by the AI agent.