Copyright © 2026 Authors retain the copyright of this article. This article is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

@article{208821,
        author = {Digvijay Vivek Phutane},
        title = {CompositeGuard: A Graph-Structural, Longitudinally-Monitored Detector for Emergent Guardrail-Bypass in Multi-Agent Systems},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {13},
        number = {4},
        pages = {2724-2732},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=208821},
        abstract = {Production guardrails for LLM-based systems check each agent action in isolation, so they cannot see unsafe outcomes that arise only when individually compliant actions are composed across agents. We call this failure class compositional guardrail leakage, give a checkable formal definition of it, and present CompositeGuard, a detector that models multi-agent execution as a typed interaction graph and flags leaking subgraphs with an exact rule-based check and a learned structural-feature classifier. We also adapt CUSUM drift monitoring, with canary calibration checks, to track the leakage rate over production time. On a 300-scenario benchmark executed through a real three-agent LangGraph pipeline, a baseline built only from local guardrails reaches 0.0 recall, while both CompositeGuard detectors reach F1 = 1.0, stable across 10 benchmark seeds. Across 20 simulated 500-run production streams in which the leakage rate rises from 2% to 18%, the monitor detected every drift with a median delay of 21.5 runs and no pre-drift false alarms. Code and benchmark are open source.},
        keywords = {Multi-Agent Systems, LLM Agents, AI Guardrails, Safety Monitoring, Drift Detection, Graph Neural Networks, Agent Orchestration, LangGraph.},
        month = {September},
        }