Consensus–Bayesian Anomaly Detection in Agentic Access Graphs

Enterprise environments increasingly include both human users and autonomous agents accessing shared software, knowledge, and infrastructure assets such as directories, repositories, and internal services. A single compromised or manipulated agent, especially under rapidly growing agentic capabilities, can propagate abnormal access across structurally coupled resources faster than existing threat-detection systems can react, making graph-level behavioral monitoring increasingly important. We model each user or agent as a node in a heterogeneous belief network over a multi-level access graph and each directory as a topic, following the consensus framework of Ye et al., 2021. Healthy access patterns correspond to opinion-dynamics regimes in which the theory guarantees per-SCC(Strongly Connected Component) consensus, whereas malicious modifications to an agent's logical dependencies break the structural-homogeneity assumption underlying those guarantees and manifest as transient rises in cross-agent opinion variance. We turn this observation into a sequential Bayesian detector, in which a binary anomaly state is updated online through an exponential likelihood on the variance jump, with both static-prior and online-prior variants. Experiments on a synthetic seven-agent/seven-directory testbed show that the detector responds to homogeneity-breaking perturbations rather than benign within-SCC(Strongly Connected Component) weight changes, that SCC-aware variance localizes the attacked block more clearly than global variance, and that online-prior recursion saturates faster than a static-prior baseline as perturbation strength grows. SCC-level decomposition yields direct structural explainability for the flagged anomaly, producing a verdict format well matched to machine-speed threat regimes and agentic security-operations workflows.