X-AgentRed: Red-Teaming Personal Agents in Evolving Multi-Service Ecosystems
X-AgentRed finds 1.77× more harmful multi-service agent paths than STAC across four frontier models.
X-AgentRed is a self-evolving red-teaming framework that discovers harmful multi-service tool-call paths for personal agents fulfilling benign requests. It builds a directed graph of each tool's effects, potential harms, and preconditions, then instantiates sampled recipes as realistic scenarios. Execution feedback diversifies harmful paths through tool substitution and repairs preconditions at blocked steps. Across four unnamed frontier agentic models, it achieves 1.77 times STAC's harmful-path rate and 2.99 times average tool coverage.
- Builds a graph from tool effects, harms, and execution preconditions
- Refines blocked paths and substitutes tools using execution feedback
- Reaches 1.77× STAC's harmful-path rate on four frontier agents
- Average tool coverage is 2.99× higher than STAC
From the original article
1 min read at arxiv.orgPersonal agents increasingly connect to services across the web and applications by using tools. As the effects of their actions accumulate and interact across services, unintended or harmful outcomes can arise when agents fulfill benign user requests. Proactively identifying these compositional failures requires systematically exploring diverse cross-service paths.
By Yeonsung Jung · 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.