arxivcs.CRcs.LG2026-07-11
NetInjectBench: Benchmarking Indirect Prompt Injection in Tool-Using Large Language Model Agents for Network Operations
Ruksat Khan Shayoni, Muhammad Faraz Shoaib, S M Asif Hossain, M. F. Mridha
Tool-using large language model (LLM) agents are attractive for network operations, but tickets, alerts, logs, runbooks, and ChatOps messages can carry indirect prompt injections. We present NetInjectBench, a 130-scenario benchmark that separates untrusted artifact text, trusted…