AgenticRepair Patches 73% of SEC-Bench Vulnerabilities

AgenticRepair patched 73% of SEC-Bench's 300 real-world vulnerabilities, beating the strongest baseline by 29 points, per a new paper posted to arXiv.

The gap between general bug repair and vulnerability repair is the forensic work security teams already do by hand. They trace data flow across files, read crash semantics, and dig through commit history before writing a fix. Automating that reconstruction, not the patch itself, pushed the success rate to 73%.

Three specialized subagents each build one piece of that context. One maps cross-file data-flow and memory-operation structure. A second reconstructs runtime crash semantics and memory origins. The commit history showing how the fragile pattern got introduced comes from the third. A fourth subagent holds all three in memory and synthesizes the patch, instead of guessing from the vulnerable line alone.

Ablations in the paper show the three context types are complementary rather than redundant: dropping any one weakens the result.

A related paper, AuditCoder, tested the opposite bet on a different benchmark. Forcing a repair agent to output an explainable construction trace cost it 7.5 to 8.5 points of pass@1. AgenticRepair spends similar effort up front and gains accuracy instead of losing it.

The paper doesn't say what's failing in the remaining 27%, or whether the three-facet approach holds up outside SEC-Bench's instance set.