ATrust: Adaptive Trust Boundary Expansion for Securing LLM Agents
Story 13, from arXiv
ATrust is a new defense mechanism designed to secure large language model (LLM) agents by dynamically expanding trust boundaries during runtime. It reduces attack success rates while improving utility compared to existing solutions.
Key points
- ATrust addresses indirect prompt injection attacks by adaptively expanding trust boundaries during agent execution.
- It starts with the user query as the root trust anchor and progressively admits task-essential information as new trust anchors.
- Evaluated on AgentDojo and AgentDyn, ATrust achieves the best security-utility trade-off among existing defenses.
- With GPT-4o as the engine, ATrust reduces attack success rate on AgentDyn from 37.80% to 4.37%.
- It outperforms AuthGraph, the state-of-the-art defense, by 12.33% in attack success rate reduction.
Why it matters
As LLM agents become more autonomous, they are increasingly vulnerable to prompt injection attacks. ATrust offers a novel approach to balancing security and functionality, which is critical for deploying these systems safely in real-world scenarios.
Counterpoint
The study is based on evaluations using GPT-4o and specific benchmarks, which may not fully represent real-world attack vectors or deployment environments.
Sources
This summary was written by a language model and may contain errors. Read the original for the full story.