A Self-Adaptive Red Teaming Framework for Vulnerability Profiling with Dynamic Attack Graphs and Retrieval Augumented Generation

The increased complexity of enterprise networks, which has been associated with microservice architectures, transient cloud infrastructure, and short software development cycles, has exceeded the capacity of conventional vulnerability management tools. Often static scanners miss threat due to banner obfuscation, while Large Language Models suffer from hallucination and an inability to reason about newly discovered vulnerabilities in red teaming. To overcome these shortcomings, this paper introduces Graph Retrieval Augmented Generation agent, a self-adaptable multi-agent red teaming architecture combining Retrieval-Augmented Generation with dynamic attack graph formation. The semantic retrieval and time relevance of our system are based on the CISA Known Exploited Vulnerabilities (KEV) catalog and the NVD 2025 Common Vulnerabilities and Exposures (CVE) feed as a hybrid vector database, which are highly precise, thereby guaranteeing accurate semantic retrieval and time relevance. Our proposed agent detects CVEs with 100 percent accuracy and retrieval latency of sub-20ms upon controlled experiments on 15 differentiated targets including obfuscated and novel services, significantly outperforming regex-based scanners and standalone standoff Large Language Models. The paper gives an explicable, scalable, and hallucination-resistant architecture of autonomous vulnerability discovery in practical security settings.