X-AgentRed
A self-evolving red-teaming framework for personal AI agents that systematically discovers harmful cross-service tool-call chains by constructing tool dependency graphs. It uses execution feedback to drive self-evolution, addressing combinatorial failure safety hazards.
Timeline (last 90 days)
X-AgentRed is a self-evolving red-teaming framework that systematically discovers harmful tool call chains through a tool dependency graph.
X-AgentRed's red-teaming agent analyzes each tool along three dimensions: effects, potential harms, and execution preconditions.
The directed graph constructed by X-AgentRed captures the dependency relationship of how one tool's effects satisfy another tool's execution preconditions.
Recipes sampled from the dependency graph by X-AgentRed are instantiated into real scenarios to test whether the agent produces harmful consequences.
X-AgentRed's self-evolution mechanism includes tool substitution diversification (deriving tool variants to replace tools on harmful paths) and precondition refinement (refining the preconditions of blocked steps on failed paths).
In tests across four frontier agent models, X-AgentRed's harmful path discovery rate reached 1.77 times that of the baseline method STAC.
X-AgentRed achieves an average tool coverage rate of 2.99 times that of baseline methods
All Facts (7)
X-AgentRed is a self-evolving red-teaming framework that systematically discovers harmful tool call chains through a tool dependency graph.
50%UnverifiedX-AgentRed's red-teaming agent analyzes each tool along three dimensions: effects, potential harms, and execution preconditions.
50%UnverifiedThe directed graph constructed by X-AgentRed captures the dependency relationship of how one tool's effects satisfy another tool's execution preconditions.
50%UnverifiedRecipes sampled from the dependency graph by X-AgentRed are instantiated into real scenarios to test whether the agent produces harmful consequences.
50%UnverifiedX-AgentRed's self-evolution mechanism includes tool substitution diversification (deriving tool variants to replace tools on harmful paths) and precondition refinement (refining the preconditions of blocked steps on failed paths).
50%UnverifiedIn tests across four frontier agent models, X-AgentRed's harmful path discovery rate reached 1.77 times that of the baseline method STAC.
50%UnverifiedX-AgentRed achieves an average tool coverage rate of 2.99 times that of baseline methods
50%