EviRepair: Multi-Agent Evidence Review Boosts Automated Vulnerability Repair

EviRepair is a multi-agent system for reliable automated vulnerability repair developed by researchers from Nanjing University and Shanghai AI Lab. The core challenge in LLM-based vulnerability repair is not code modification capability but accurately understanding the vulnerability origin, attack path, and required preservation of normal behavior.

Two-Phase Architecture: Diagnosis Then Synthesis

The repair process splits into two phases connected by a reviewed "vulnerability model":

Phase 1: Vulnerability Diagnosis

Instead of treating the vulnerability report as ground truth, an Analysis Coordinator agent reads the report and repository, proposes entry points, dangerous operations, and vulnerability types, then dispatches three specialized agents in parallel:

Reachability Analysis Agent : Verifies whether attacker-controlled inputs can actually reach dangerous operations along code paths.

Vulnerability Reproduction Agent : Builds an executable reproducer to observe pre-fix behavior and defines runtime conditions that a correct fix must satisfy.

Boundary Analysis Agent : Determines which inputs must be blocked and which legitimate inputs and existing behaviors must be preserved.

A Reviewer Agent then checks the evidence from these three agents for consistency, producing a verified vulnerability model containing five elements: root cause, attack path, runtime decision conditions, security boundaries, and required preserved behaviors.

Phase 2: Patch Synthesis

Using the verified vulnerability model, a Patch Generator agent creates candidate fixes. An independent Reviewer Agent validates each patch against the same vulnerability model, ensuring it satisfies security boundaries and compatibility requirements. This dual-review design catches issues missed in the first phase.

Experimental Evaluation

Experiments on two benchmarks:

PatchEval : 230 vulnerability repair tasks covering Go, Python, and JavaScript.

SEC-bench : 300 real-world security tasks including CVEs and OSS-Fuzz issues.

Systems compared: EviRepair, Codex, and OpenCode, each backed by three different base LLMs. Primary metric: Resolved Rate (proportion of patches passing hidden tests).

Results show EviRepair outperforms baselines across all base models on both benchmarks. The advantage is especially pronounced on SEC-bench CVE subset where vulnerability reports often fail to identify the true root cause, requiring the system to reconstruct attack paths and security boundaries—exactly what the evidence-driven diagnosis phase provides. On OSS-Fuzz tasks with explicit crash information, the gap narrows.

Ablation Study: Both Review Stages Are Essential

Replacing either the diagnosis-phase review or the synthesis-phase review with a single-agent process reduces performance. Replacing both with an end-to-end single-agent pipeline degrades further. This confirms that the gains come from structured evidence review, not merely from additional model calls. The diagnosis review reduces incorrect vulnerability understanding; the synthesis review ensures fixes truly meet security and compatibility constraints. The two reviews complement each other, allowing issues missed in the first phase to be caught in the second.

EviRepair: Evidence-Driven Agents for Reliable Vulnerability Repair
https://arxiv.org/abs/2610.13491
https://github.com/lly-s/EviRepair

Figures and tables from the paper illustrate the workflow comparison, benchmark results, and ablation data.