Semantic Drift
LLMs often propose "hallucinated" fixes in the wrong files or break existing business logic.
False Safety
Models confidently claim a vulnerability is fixed, but the fix is only superficial (e.g., adding a comment or a useless check).
We need a system that moves from "Generation" to "Execution-Verified Mitigation."
General coding agents (SWE-agent) optimize for passing unit tests.
Automated Vulnerability Repair (AVR) must optimize for eliminating sinks.
Structural Matching: Penalizes innovative, valid patches.
LLM-as-Judge: Prone to bias and lacks "executable proof," leading to a false sense of security.
Unlike C/C++ (where crashes/segfaults provide clear signals), web vulnerabilities are often "Silent Failures", which means the app runs fine, but remains exploitable.
Security repair is a high-stakes, multi-step process. A single prompt cannot effectively manage file discovery, root-cause analysis, and regression testing simultaneously.
Single agents often suffer from "confirmation bias", in which they believe their own patch works without external verification.
Based on these insights, we propose a Multi-Agent Workflow that decomposes the complex repair task into specialized roles, ensuring higher precision and reliability.
Issue Review
Analyze security-related issue reports and produce structured findings.
Pull Request Review
Review code changes in PRs and provide risk analysis with mitigation suggestions.
Repo-wide Full Scan
Run baseline security scanning across the whole repository.
Repo-wide Incremental Scan
Re-scan only the changed scope between base/head for continuous security assurance.
Methodology is shared across workflows; here we use repo-wide scan as the most complete example.
Systematic scanning of the entire codebase or specific change sets (Diffs).
Identification of potential vulnerability candidates and security hotspots.
The analyzer audits input signals together with repository evidence to decide whether a case is confirmed.
Analyzer Gate: If the case does not pass this gate, workflow stops before mitigation; if it passes, the agent verifies root cause, models exploit chains, and assigns dynamic CVSS v4 scores.
Mitigate: Generates Minimal Patches that remove vulnerability sinks while preserving business logic.
Verify: Audits original input, analyzer conclusions, and patch content together with regression checks.
Verifier Gate: If verification does not pass, mitigation output is blocked while validated analysis remains available for reporting.
Correlates related fixes and consolidates them into unified delivery bundles.
Produces traceable artifacts for evidence, reasoning, and patch status, then automates report and GitHub Pull Request generation under explicit human approval.
To move beyond prompt-only reasoning, each stage operates on grounded repository context.
Merge Before Analyzer (Reduce Analysis Load)
Merge After Verifier (Reduce Reviewer Overload)
CVSS v4 Scoring (Prioritize Reviewer Attention)
High-Fidelity Samples: Each test case serves as an independent "miniature repository" representing real-world vulnerability scenarios.
Triad Data Structure:
Scenario Simulation
Transforms each PatchEval case into a dedicated Agent Workspace.
End-to-End Mitigation
The Agent autonomously performs path discovery, analysis, and patch generation based on the provided descriptions.
Execution-Based Verification
System automatically executes accompanying test scripts; a fix is marked successful only if all tests pass without regressing original functionalities.
Real-world Vulnerability Discovery & Mitigation in a Simulated E-commerce Environment
Long-term Defensive Utility: Similar to the evolution of Fuzzers, autonomous agents are expected to become essential defensive pillars. While they accelerate discovery, their long-term value lies in empowering defenders through automation.
The Case for Staged Deployment: To mitigate potential short-term shifts in the threat landscape, these systems should initially be limited to organizations with robust governance and auditing capabilities, ensuring a responsible path to public release.
Contribution
An agentic AVR framework with explicit gates and closed-loop verification.
Evaluation
PatchEval benchmark · Vulnerable-Book-Shop case study.
Ethical Stance
Blue-team only · sandboxed · staged release.