ATrust adaptively expands trust boundary to secure LLM agents
Original title: ATrust: Adaptive Trust Boundary Expansion for Securing LLM Agents
ATrust reduces AgentDyn attack success rate from 37.80% to 4.37% while boosting utility by 10% with GPT-4o.
ATrust begins at the user query as the root trust anchor and progressively admits observations based on execution state. It evaluates external content against the current trust boundary to exclude irrelevant or malicious input. Evaluation on AgentDojo and AgentDyn demonstrates a superior security-utility trade-off compared to existing defenses.
- Adaptive trust boundary expansion during runtime
- Starting from user query as root trust anchor
- Reduces attack success rate on AgentDyn to 4.37%
- Improves utility by 10% over baseline
- Outperforms AuthGraph by 12.33% in security-utility balance
From the original article
1 min read at arxiv.orgAs large language model (LLM) agents take on more complex tasks with greater autonomy, security has become a major obstacle to their broader deployment. Unlike in attacks that require complex exploitation of operating system vulnerabilities, attackers can easily hijack agents through indirect prompt injection (IPI), leading to privacy leakage, financial loss, and destructive operations.
By Xu Zhang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.