X-AgentRed: Red-Teaming Personal Agents in Evolving Multi-Service Ecosystems
The paper proposes a feedback-driven way to find harmful cross-service tool paths that ordinary prompt-based red-teaming may miss.
X-AgentRed builds a dependency graph from tool effects, harms, and execution preconditions, then samples paths that can turn benign agent requests into harmful outcomes. Failed runs are used to refine blocked preconditions, while successful harmful paths are diversified through tool substitution. In tests across four frontier agentic models, it reports 1.77x the harmful-path rate of STAC and 2.99x average tool coverage. Source: ArXiv · AI/CL/LG's note.
X-AgentRed builds a dependency graph from tool effects, harms, and execution preconditions, then samples paths that can turn benign agent requests into harmful outcomes. Failed runs are used to refine blocked preconditions, while successful harmful paths are diversified through tool substitution. In tests across four frontier agentic models, it reports 1.77x the harmful-path rate of STAC and 2.99x average tool coverage. Source: ArXiv · AI/CL/LG's note.
score 5