A new research framework makes AI agents prove a security bug is real before letting them patch it.
The system, described in a paper on arXiv, tackles vulnerability detection and repair across Java, Python, and C++ by converting all three into a shared structural format called a Universal Abstract Syntax Tree. It pairs a graph-based model (GraphSAGE) with code embeddings from Qwen2.5-Coder-1.5B, then runs detection, execution-based validation, and repair as three distinct stages. The hard rule: no automated fix happens until the tool actually confirms the flaw is exploitable, not just statistically likely. In testing, it hit 89.84-92.02% accuracy spotting bugs within a single language, 74.43-80.12% F1 on languages it wasn't trained on, and resolved 69.74% of confirmed vulnerabilities end to end, with a 12.27% overall failure rate.
This matters because AI coding agents are quietly being handed write access to real codebases, and most of them act on a classifier's best guess rather than checked evidence. The paper's own ablation makes the stakes concrete: turning off the validation step caused unnecessary repairs to jump 131.7%, which is the agentic-AI equivalent of a doctor operating on a hunch. That's the same failure mode showing up in AI-assisted coding generally - plausible-sounding output standing in for verified fact.
Still, a 12% failure rate and an 80% ceiling on unfamiliar languages means this is a meaningful step, not a solved problem, and arXiv self-reported benchmarks deserve the usual grain of salt until independently replicated.