PromptShield AI: A Multi-Agent Architecture for Intelligent Prompt Injection and Jailbreak Attack Detection Using Machine Learning
Jahnavi Somaraju, N. Sree Charan, M. Mythili, T. Reddy Bhargavi, K. Navya Sree
Large language models (LLMs) are increasingly deployed in user-facing applications, which exposes them to prompt injection and jailbreak attacks that override system instructions, exfiltrate data, or elicit disallowed behaviour. Existing defences are largely single-mechanism: a rule filter, a single machine-learning (ML) classifier, or a single LLM-based judge, each of which is comparatively easy to evade once its decision boundary is known. This paper proposes PromptShield AI, a multi-agent architecture that fuses lexical, statistical, semantic, and contextual detection signals through a coordinated set of specialised agents rather than a single monolithic classifier. An Orchestrator Agent decomposes each incoming request and routes it to five parallel detection agents; a Decision Fusion Agent calibrates and combines their outputs; a Response/Mitigation Agent executes the resulting policy (allow, sanitize, or block); and a Logging and Feedback Agent closes the loop for continuous retraining. We describe the architecture, the inter-agent communication protocol, the fusion algorithm, and a reference implementation, and we outline an experimental protocol together with illustrative evaluation results and an ablation study. The results indicate that the coordinated multi-agent design improves detection F1-score and reduces false-positive rate relative to any individual detection mechanism, at an acceptable latency overhead, while remaining more resilient to obfuscation and multi-turn attacks than single-agent baselines.