Cybersecurity training struggles to keep pace with the rapid emergence of new vulnerabilities and attack techniques. Traditional methods rely heavily on manual setup, static exercises, and delayed integration of real-world threats, leading to outdated and ineffective learning experiences. Meanwhile, the accessibility of Large Language Models (LLMs) has enabled attackers to automate vulnerability discovery and exploit generation, emphasizing the urgent need for adaptive and continuously updated defensive training tools. This paper proposes HackEd, an automated system that uses LLMs and Retrieval-Augmented Generation to create and personalize Capture-the-Flag challenges. The system operates in two phases: challenge creation, where data from multiple knowledge bases is processed to generate vulnerability-based exercises packaged as Docker containers, and experience tailoring, where an adaptive LLM tutor monitors user performance and provides real-time hints and feedback according to each trainee’s skill level. A human-in-the-loop evaluation was conducted to assess technical validity and pedagogical value. The generated challenges were reviewed for realism and educational effectiveness, while controlled student experiments measured engagement and learning outcomes. Results show that LLMs, when combined with structured knowledge bases, can generate relevant challenges with minimal human intervention. Moreover, the adaptive tutor significantly improved learners’ comprehension, motivation, and retention by offering individualized guidance during problem-solving. Technically, the system reduces the cost and time required to update cybersecurity training content. Pedagogically, it enhances engagement and inclusivity by dynamically adjusting difficulty and feedback. Overall, HackEd demonstrates the potential of integrating LLM-driven automation into cybersecurity education, providing a scalable, efficient, and continuously evolving framework to prepare the next generation of cybersecurity professionals.