Meta researchers map cross-service risks in personal agents
X-AgentRed uses tool-dependency graphs and execution feedback to search for potentially harmful workflows.
Event / data period: 2026-10-09
Published: · Updated:
Meta researchers introduced X-AgentRed, a red-teaming framework for studying risks that can emerge when personal AI agents use tools across multiple services. It builds a dependency graph showing how one tool’s effects may satisfy another’s prerequisites, then expands or repairs candidate scenarios using execution feedback.
Across four frontier agent models, the authors report an average harmful-path rate 1.77 times the STAC baseline and tool coverage 2.99 times as high. These ratios were measured in the researchers’ constructed scenarios and simulated environment; they are not rates of incidents involving real users. The paper also notes the computational cost of updating the dependency graph as services change.
Data
Measure
Evaluation in the researchers’ constructed red-team scenarios
| Number of agent models tested | 4 | Evaluation in the researchers’ constructed red-team scenarios |
Methodology
Values come from the paper's stated benchmark experiments; they are not field measurements or general performance guarantees.
Updated:
Subjects and places
Subject proposals
- Science and technologySuggested
- Preserved editorial category/topic; broad candidate onlyarxiv.org ↗
Geographic references
Only source-provided geographic references are shown. Opening an external map is optional. Optional story maps use OpenStreetMap. Opening a map sends your IP address and this site’s address to the map provider; location data remains readable as text without loading the map.
- United States · Place role: Coverage · Place type: Country or area · Location precision: RegionalEvidence
- Explicit preserved editorial country/region; no city or national centroid inferredarxiv.org ↗