BAg-IDS: A Blockchain-Enabled Agentic AI Architecture for Trustworthy Intrusion Detection

Abstract Learning-based intrusion detection produces verdicts operators must simply trust: the audit record is mutable and no independent party endorses the decision. We present BAg-IDS, a distributed detection architecture in which autonomous agents detect, independently re-check, and co-sign every remediation-triggering verdict on a Hyperledger Fabric ledger before it is acted upon, placing certification on the decision path rather than in an after-the-fact log. A calibrated, reliability-weighted ensemble with a reject option feeds the quorum; language-model and governed human-in-the-loop tiers handle escalation. The full stack is implemented and every number measured. Selective certification raises accuracy from 96.6% to 99.7% (0.04% false-positive rate) at 87.3% coverage on CICIDS2017, and reaches 99.98% on the independent GeNIS 2025 corpus while deferring 0.14% of flows. A 360-alert benchmark shows the ~ 2 s consensus latency is structural: one write per key per block forces K votes across at least K blocks, a bound no submission strategy removes.