Neural Sentinel: Automated Security Auditing for LLM, RAG, and Agentic Systems Under the EU AI Act

The EU AI Act (Article 15) mandates robustness and cybersecurity for high‑risk AI systems. Small and medium‑sized enterprises(SMEs) using Large Language Models (LLMs), Retrieval‑Augmented Generation (RAG), or agentic workflows lack affordable toolsto assess compliance. This paper introduces Neural Sentinel, a suite of five gRPC‑based security scanners that automate the detectionof critical vulnerabilities: prompt injection, system prompt leakage, data exfiltration, document poisoning, tool hijacking, memorypoisoning, and post‑quantum cryptography weaknesses.Unlike open‑source scanners (Garak, Counterfit), Neural Sentinel produces a 1‑page compliance report mapped directly to the EU AIAct. Unlike commercial scanners (Lakera, ProtectAI), it offers a free, asynchronous audit model requiring only the client’s systemprompt – no API key, no integration. Attack taxonomies (RAG01‑07, AGT01‑08) are documented, and eight public demo videosdemonstrate real critical findings. The scanners are live on the SingularityNET mainnet (Ethereum) and have been tested against100+ synthetic prompts, achieving an 86% detection rate for known injection patterns. This paper demonstrates that low‑cost,automated, compliance‑oriented security audits are feasible and can help Italian SMEs meet the August 2026 EU AI Act deadline.