CyberForge: Verified Vulnerability Injection at Repository Level for Cybersecurity Agent Training

2026-08-06

AI Summary

This work addresses the scarcity of scalable, verifiable warehouse-scale security training data for defensive AI agents by proposing a novel approach that does not rely on historical CVEs. Leveraging program analysis and automated vulnerability injection, the method generates executable and dynamically validated vulnerability samples within real-world C/C++ projects: the injected variants pass unit tests but trigger proof-of-vulnerability (PoV) exploits only when the vulnerable code is present. The resulting dataset comprises 1,034 verified vulnerabilities across 80 projects and 63 weakness types, exhibiting locality comparable to real CVE patches. Models fine-tuned on this data achieve gains of 3.3–14.7 points on SEC-bench repair tasks and outperform teacher models on cross-language PatchEval, demonstrating strong generalization capabilities.

Abstract

Despite recent advances, frontier large language model (LLM) agents remain limited in discovering and patching complex vulnerabilities in real-world software. Generally available agents can already aid attackers, who only need to find one exploitable weakness, while defenders must continuously identify and patch all vulnerabilities across fast-growing codebases. Stronger defensive agents would help close this gap, yet the scarcity of security training data with reproducible build and execution environments remains a bottleneck. We present CyberForge, a framework that synthesizes executable, repository-level security training data by injecting vulnerabilities into real C/C++ projects. It validates each instance dynamically: the injected build must pass the project's unit tests, and generated proof-of-vulnerability (PoV) must trigger on the injected build and not on the clean one. CyberForge is not limited by the availability of disclosed vulnerabilities, therefore it can scale in comparison to data augmentation techniques which rely on historic CVE data. The resulting corpus holds 1034 validated vulnerabilities across 80 projects and 63 weakness categories, with edit locality similar to real CVE patches under a real-versus-real noise floor. Fine-tuning on trajectories collected over this corpus improves SEC-bench patch repair by +3.3 to +14.7 points, in all six configurations of three model scales and two teachers, with the 31B student reaching its GPT-5.4-mini teacher, 72.7% against 74.0%. These gains generalize out of distribution to PatchEval, a corpus containing other programming languages, where every configuration also improves and the 31B student passes its teacher.

Problem

• vulnerability injection
• security training data
• cybersecurity agent
• repository-level
• defensive AI

Innovation

• vulnerability injection
• repository-level synthesis
• proof-of-vulnerability
• executable training data
• LLM agent training