AgenticRepair Hits 73% on SEC-Bench by Engineering Three Context Facets Security Engineers Assemble by Hand

arXiv 2607.29422·high signalXBlueskyLinkedInCopy link

The authors identify what agentic vulnerability repair is missing versus general bug repair: cross-file data-flow and memory-operation structure, runtime crash semantics and memory origins, and commit history showing how the fragile pattern was introduced. AgenticRepair runs three specialized LLM subagents to build those contexts, then embeds them in a dedicated repair subagent's memory for context-conditioned patch synthesis. On SEC-Bench's 300 real-world instances with sanitizer-based patch verification it reaches a 73% success rate, outperforming the strongest baseline by 29%, and ablations confirm the three facets are mutually complementary rather than redundant.