Decentralized Agentic AI Security Framework Using Blockchain and Reinforcement Learning with Explainable AI

The rapid evolution of cyber threats targeting AIdriven systems necessitates robust, adaptive security frameworks. This paper presents Agentic AI Security Platform, a novel architecture integrating deep reinforcement learning (DRL), distributed watchdog monitoring, and blockchain-based auditing to detect and mitigate multi-modal cyber threats. The system employs a Deep Q-Network (DQN) model trained on heterogeneous data streams-textual, visual, and auditory inputs-to classify threats such as Data Exfiltration, advanced persistent threats, DDoS Attacks, Phishing campaign, zero-day exploits and supply chain attacks with 98.2 % accuracy, reducing false positives by 32 % compared to rule-based baselines. A distributed watchdog mechanism augments detection by monitoring network traffic in real time, achieving 95 % anomaly detection rates with 2.1 -second latency, while blockchain technology ensures tamper-proof logging of security events via Ethereum smart contracts, demonstrating 94.6% transaction success rates on the Sepolia test net. The framework incorporates SHAP (SHapley Additive explanations) values to provide interpretable threat classifications, addressing the “black box” challenge in AI security. Experimental validation across $\mathbf{1 2, 0 0 0}$ adversarial samples and $\mathbf{1, 8 2 4}$ blockchain transactions reveals a 14 % improvement in detection accuracy over state-of-the-art systems, with sub-second response times $(327 ~\text{ms}$ avg). The platform's modular design enables seamless integration into existing AI infrastructures while balancing security and usability through dynamic, risk-proportional countermeasures. This work advances secure AI deployment in critical applications by unifying adaptive threat detection, explainable decision-making, and cryptographic auditability.