用多维度代码上下文提升AI修复漏洞的准确率

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.