CyberForge injects validated C/C++ flaws to train defensive security agents

Frontier large language model agents still struggle to find and patch complex vulnerabilities in real software, an arXiv preprint says, while available agents already help attackers who need only one exploit. Stronger defensive agents are limited by scarce security training data with reproducible builds and execution environments. The authors present CyberForge, a framework that synthesizes executable, repository-level security training data by injecting vulnerabilities into real C and C++ projects. Each instance is validated dynamically: the injected build must pass the project's unit tests, and a generated proof-of-vulnerability must fire on the injected build and not on the clean one. Because the method is not limited to disclosed flaws, it can scale beyond data augmentation that depends on historic CVE records. The resulting corpus contains 1,034 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 the corpus improved SEC-bench patch repair by 3.3 to 14.7 points across six configurations spanning three model scales and two teachers. A 31-billion-parameter student reached 72.7% against its GPT-5.4-mini teacher's 74.0%. Gains also held out of distribution on PatchEval, which includes other programming languages, where every configuration improved and the 31B student passed its teacher.