AgenticRepair: Multi-Faceted Program Context Engineering for Agentic Vulnerability Repair
Michael Fu, Qiyue Mei, Patanamon Thongtanunam, Kla Tantithamthavorn
Abstract
Automated vulnerability repair aims to reduce the time and effort required to patch security flaws from a vulnerability triage report. Recent agentic AI approaches have shown promising results in automated program repair. However, vulnerability repair demands richer program context than general bug repair - context that security engineers routinely assemble in practice but that existing agentic approaches do not engineer. We identify three critical gaps: code-structure context capturing cross-file data flows and memory operation patterns, runtime-execution context revealing crash semantics and memory origins, and commit-history context recovering how fragile code patterns were introduced. We present AgenticRepair, an agentic vulnerability repair framework that addresses the gaps through multi-faceted program context engineering. AgenticRepair orchestrates three specialized LLM subagents to engineer the contexts, which are then embedded into the memory of a dedicated repair subagent for context-conditioned patch synthesis. Evaluated on SEC-Bench comprising 300 real-world instances with sanitizer-based patch verification, AgenticRepair achieves a 73% success rate, substantially outperforming the strongest baseline by 29%. Our ablation study confirms that the three context facets are mutually complementary, and that multi-agent scaffolding and base-model capacity each play an essential role. Collectively, these findings establish multi-faceted program context engineering as a promising design direction for agentic vulnerability repair.
Topics
Classified with taxonomy v2 on Sat, 5 Sept 2026.