An agent can fulfill a legitimate request while causing damage elsewhere. X-AgentRed explores the tool chains and tests them by execution.
The security of a personal assistant also depends on the interactions between its services. X-AgentRed builds a graph where the effects of one tool can make it possible to use another, and then transforms these paths into test scenarios.
The system leverages runtime results to diversify its testing and correct stalled steps. Across four agent models, the authors measure 1.77 times more harmful paths than STAC and an average tool coverage 2.99 times wider. The study focuses on compound failure detection, not a production safety guarantee.
Preprint filed on arXiv on October 9, 2026; results are those reported by the authors and remain to be independently confirmed.
Read the full article on the original source
Go to the source →