Prompt Injection Resistance in Tool-using Llm Applications: Architectural Patterns and Failure Modes
Type: article
Abstract: Tool-using large language model applications combine language generation with retrieval, messaging, database access, code execution, and other actions. Prompt injection therefore becomes a security-boundary problem: untrusted content can influence decisions made with application or user privileges. Model alignment, prompt structuring, and detection reduce attack success in bounded settings but remain vulnerable to distribution shift, adaptive payloads, and dynamic planning. Stronger resistance comes from separating control and data flows, issuing task-scoped capabilities, mediating every tool call, validating structured actions independently, isolating memory, and preserving provenance. A reference architecture and evaluation framework are proposed. Prompt injection should be treated primarily as an authorization and information-flow problem rather than as a wording defect.
Source: CyberLeninK (CyberLeninka)
Author: Zakharia Andrey
Institution: University of Belgrade
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7168066883