Hijacking Tool-Integrated LLM Agents via Indirect Prompt Injection: A Benchmark Scaffold and Threat Taxonomy
Abstract
Tool-integrated large language model (LLM) agents compose privileged actions — file reads, email dispatch, shell invocation — from the reasoning traces of a probabilistic model. Because the same channel carries both task data and latent instruction, any adversary who can place text into the data pipeline gains a tool-call influence that is bounded only by the agent's authorization envelope. Indirect prompt injection exploits this conflation. The security properties of such agents are not yet well characterized in the open literature: reports exist, but reproducible infrastructure, statistically grounded protocols, and explicit threat decompositions remain scarce. This paper does not claim a novel attack. It contributes three artifacts that the prior reports lack: (i) a reproducible benchmark scaffold for indirect prompt injection against a LangChain-style ReAct agent, parameterized over model, tool set, and injection medium; (ii) a consolidated threat taxonomy that unifies the HTML-comment, Base64, and hidden-CSS vectors introduced by Greshake et al. [1] with a supervisor-reframing adaptive variant designed to defeat system-prompt defenses in the sense of Perez and Ribeiro [2]; and (iii) an evaluation protocol specifying trial counts, confidence intervals, position-sensitivity ablations, and detection-latency measurements that a downstream empirical study can execute against real Llama 3 8B and 70B endpoints [3]. The scaffold is designed to be composable with InjecAgent [4] and AgentDojo [5]. We explicitly decline to report attack-success rates in this paper: the full empirical pass against live model endpoints is the next release of the artifact. What we do provide is the apparatus, the power analysis, and the qualitative behavioral shape of representative ReAct traces collected during scaffold development, clearly labeled as such. The contribution is infrastructural, not declarative.