Adversarial Vulnerabilities in Large Language Models: A Critical Review and an Adaptive Multi-Layered Defense Framework for Enterprise AI

Large Language Models (LLMs) now sit at the centre of many digital services, supporting the generation, summarisation, interpretation, and use of natural-language and multimodal information. As these models are connected to retrieval-augmented generation (RAG), external tools, cloud platforms, enterprise knowledge bases, and autonomous agents, the security boundary extends well beyond the model itself. This article critically reviews adversarial vulnerabilities across the wider LLM ecosystem using peer-reviewed studies, influential preprints, and authoritative standards available up to July 2026. It examines prompt injection, jailbreaks, data and model poisoning, backdoors, model extraction, membership inference, training-data extraction, RAG manipulation, supply-chain compromise, multimodal attacks, and the misuse of agentic tools. Current safeguards, including reinforcement learning from human feedback, adversarial training, prompt and output filtering, retrieval validation, differential privacy, red teaming, and explainability, reduce some risks but remain inadequate against adaptive and multi-stage attacks. In response, the article introduces the Adaptive Multi-Layered Adversarial Defense Framework (AMADF), a ten-layer enterprise architecture covering identity and access management, prompt sanitisation, retrieval provenance, semantic threat detection, secure inference, explainability-based risk scoring, output validation, human review, continuous monitoring, and governance. AMADF draws on Defense-in-Depth, Zero Trust Architecture, the NIST AI Risk Management Framework, Explainable AI, and socio-technical security principles. It offers organisations a practical structure for designing, assessing, and governing secure LLM deployments, although its effectiveness still requires empirical testing in operational environments.

Keywords: Large Language Models, adversarial machine learning, prompt injection, jailbreak attacks, retrieval-augmented generation, AI agents, Zero Trust, Explainable AI, trustworthy artificial intelligence, cybersecurity.