Architecting Secure AI Agents: Perspectives on System-Level Defenses Against Indirect Prompt Injection Attacks

2026

Open publication workspace · Sign in to read the full PDF.

AI-generated summary

1) This paper proposes a system-level defense architecture to secure AI agents against indirect prompt injection attacks by focusing on dynamic replanning, constrained model usage, and human interaction.
2) * Outlines three core positions for system-level defenses: dynamic replanning and policy updates, the necessity of LLMs for certain security decisions, and the importance of personalization and human interaction in ambiguous cases.
* Discusses limitations of current benchmarks that create a false sense of security and utility.
* Proposes concrete strategies for using LLMs in security decisions, emphasizing constrained inputs and tasks to mitigate prompt injection risks.
3) AI Agents, Security, Prompt Injection, System-Level Defenses, LLM Security

Check the original publication for accuracy and context.