As autonomous AI agents transition from sandboxed experiments to enterprise production environments, traditional reactive and identity-only security models have proven insufficient. These models fail to address the real-time, path-dependent nature of autonomous decision-making and the emerging threats of prompt injection, goal substitution, and semantic drift. This paper presents a novel runtime governance architecture comprising a 10-layer defence-in-depth enforcement pipeline. The framework introduces path-dependent policy evaluation, semantic intent alignment, and mandatory Client-Initiated Backchannel Authentication (CIBA) for high-risk actions. Furthermore, we detail the integration of post-quantum cryptographic identity management (hybrid Ed25519 and ML-DSA-65 / FIPS~204) across both server-side gateways and two distinct mobile companion applications: the Aegis CIBA companion (enterprise operator HITL approval) and the Sentinel companion (vendor-level biometric-gated PQC authentication). Empirical evaluation of a preproduction reference implementation comprising 1, 098 automated tests demonstrates that the structural and semantic perimeter introduces sub-millisecond per-layer overhead (<0. 25\, ms). Through rigorous load testing, we identify a PostgreSQL connection pool bottleneck at 150\, req/sec (p95 latency 1, 610\, ms) and demonstrate that a Redis-backed read-through cache optimisation yields an 85 improvement, reducing p95 latency to 19\, ms while maintaining a 100\% detection rate across the evaluated adversarial test corpus for Fine-Grained Authorisation (FGA) violations under sustained load. The architecture natively supports publicly routable gateway endpoints, enabling seamless cellular CIBA device pairing over Wide Area Networks (WAN).
Patrick Matthews (Mon,) studied this question.