Arxiv AI Exploit - Agent Executing Prompt-Injected Shell (T1059.004)

Ibrahim Saud@tektrix

•updated Jun 20, 2026•0•0•2

Identifies Python-based LLM agents (LangChain, etc.) blindly passing unsanitized, prompt-injected strings to native command shells (cmd, bash). This rule specifically looks for Python processes initiating command shell executions where the command line of the shell contains common data exfiltration or remote access tools like curl, wget, nc, or base64.