GLITCH_AI: A Hybrid Framework for Automated Penetration Testing with LLM-Driven Adaptation and Reporting

Modern attack surfaces and restricted execution environments make manual penetration testing slow and difficult to scale. This paper presents Glitch_AI, a hybrid framework that orchestrates established security tools with LLM-assisted error recovery and reporting. Glitch_AI employs a FastAPI backend, a web interface, and role-based execution to support both local and restricted settings via remote fallback mechanisms. Large language models are used for guided error recovery, result interpretation, and report generation, while execution remains tool-driven and fully logged for auditability. Evaluation on intentionally vulnerable targets shows improved workflow continuity through structured retries and fallback execution, while producing auditable logs and readable reports for both technical and non-technical audiences.