This research introduces a novel approach for automated vulnerability repair using Large Language Models with dynamic state guidance, addressing the challenge of unpatched security flaws in software systems.

• Employs a state-guided debugging agent that reasons through vulnerability repair with real-time execution feedback
• Demonstrates superior performance compared to existing LLM-based repair methods across various vulnerability types
• Reduces manual security patching effort while maintaining high-quality fixes
• Introduces new benchmarks for evaluating automated vulnerability repair capabilities

This advancement matters for security teams by significantly reducing the time between vulnerability discovery and patching, ultimately minimizing exposure to potential attacks and strengthening overall system security posture.

Agent That Debugs: Dynamic State-Guided Vulnerability Repair