Provenance-gated tool invocation for defending LLM agents against indirect prompt injection
Year: 2026
Type: article
Abstract: Indirect prompt injection allows the tool-calling agent to execute calls on untrusted data, and the tool whitelist does not review the source of parameters. This paper introduces provenance gate (PGTI), which sequentially performs tool whitelist and parameter source taint checks before the tool execution, allowing the model to still see all tools. In four AgentDojo tasks, 100 attack pairs, and 60 benign tasks, Qwen2.5 14B ran 960 times. The attack success rate decreased from 17.0% with no defense to 3.0%, and it still held after switching models and attack templates; the utility under attack decreased from 40.0% to 28.0%. The benign utility relative to static tool filtering (35.0% and 26.7%) and the increment over the whitelist alone showed no significant difference in this sample. Pair comparison simultaneously reported both safety and utility.
Authors: Lulu Jing(Conceptualization, Formal analysis, Methodology, Software, Writing – original draft, Writing – review & editing), Nanmei Chen(Data curation, Investigation, Writing – review & editing), Zhongmiao Sun(Validation, Visualization, Writing – review & editing), Chao He(Resources, Validation, Writing – review & editing), Fan Si(Project administration, Supervision, Writing – review & editing)
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7233938917
