With over 25,000 new CVEs disclosed annually, a significant gap exists between vulnerability intelligence and practical training content for cyber range environments. Manual scenario development requires substantial expert effort and cannot keep pace with emerging threats. This paper presents a three-agent Large Language Model framework that automatically generates structured cyber range scenarios from public vulnerability data. The system employs a vulnerability analyzer that enriches CVE data from the National Vulnerability Database and ExploitDB, a scenario architect that maps vulnerabilities to MITRE ATT&CK techniques and constructs multi-step attack flows, and a deployment generator that produces Ansible playbooks and Docker/Podman Compose files for infrastructure-agnostic deployment. We evaluate the framework on 28 high-severity CVEs across three open-source models (Mistral-Small 3.1, Qwen3.5, Llama 3.3) and one proprietary baseline (Claude Sonnet), showing that LLM-augmented generation substantially increases ATT&CK technique diversity and multi-host topology realism over rule-based templates while producing structurally valid deployment artifacts across all configurations.