Indirect Prompt Injection in Agentic Systems: Evidence from Tool Outputs and Third-Party Content
When the Environment Talks Back: Data, Instructions, and the Duty to Disregard
Abstract
Large language model agents differ from chatbots in one decisive respect: they act. They read files, call tools, send messages, and execute commands, which means the content they ingest can also instruct them. Indirect prompt injection — instructions embedded in web pages, files, tool outputs, and messages from other agents — is now ranked by OWASP as the leading risk for LLM applications, and benchmarks show that a substantial share of tool-integrated agents follow injected instructions. This article examines how a deployed personal AI agent resists such injection in everyday work, through a single-case study of Kesang, a personal conversational agent used intensively in Indonesia between October and November 2026. Drawing on interaction records, system logs, and scheduled-task reports, the study reconstructs four episodes: a tool result displaying the contents of a maintenance script; a background-task report that failed a prompt-injection check; a third-party persona prompt arriving through the request path; and a standing pattern of retrieved content attempting to redirect the agent’s task. The findings show that effective defence is architectural — provenance tracking, an instruction hierarchy, and least-privilege tooling — rather than a matter of judging each injected message on its merits, and that a refused injection should be disclosed, not silently dropped, so that it leaves an auditable trace.
References
Downloads
Published
Issue
Section
License
This work is licensed under a Creative Commons Attribution 4.0 International (CC BY 4.0) license.