Cyber-defence mechanisms for artificial intelligence to everything (AI2X) and security to everything (SEC2X) are required to learn across many domains without centralising sensitive traffic. Although this requirement is supported by federated intrusion detection, performance can be affected by non-IID traffic, poisoned updates, backdoor manipulation, unreliable clients and deadline-bound response constraints. In this paper, Q-AFedSec is presented as an AI2X-SEC2X framework in which BDIx agentic reasoning, QUBO-based orchestration, robust federated aggregation, validation-gated commitment, AI/ML belief fusion and post-quantum-secured update exchange are combined. Local intrusion-detection evidence is maintained by each client, while beliefs about trust, drift, poisoning likelihood, latency, privacy exposure and expected detection contribution are constructed by the coordinator. Clients, defences, aggregation mode and response commitment are selected by a binary optimisation layer under trust, budget and deadline constraints. High-trust ensemble, prototype-distance and nonlinear local-model beliefs are fused by the detector. A MATLAB evaluation over binary NSL-KDD was conducted using 12 clean, non-IID, poisoning, backdoor, drift, many-client, low-participation and privacy-noise cases. Four representative approaches were used for comparison: FedAvg, FedProx, Krum and FLTrust-inspired TrustWeighted aggregation. Across the 12 cases, average accuracy and Macro-F1 of 99.41%, an average false-alarm rate of 0.18% and an average attack-success rate of 1.01% were achieved by Q-AFedSec. The findings indicate that robust federated security decisions can be improved by AI2X-SEC2X agentic belief fusion compared with fixed aggregation or trust-only baselines.

AI2X, SEC2X, artificial intelligence to everything, security to everything, federated learning, intrusion detection, BDIx agents, QUBO, quantum-ready QUBO orchestration, post-quantum security.