# Preprint tests execution checks before AI repairs software vulnerabilities

_Published Monday, October 5, 2026 at 2:06 AM EDT · Science, AI · Latest · Tier 2 — Notable_

Researchers report in a preprint that an AI framework resolved 69.74% of software vulnerabilities end-to-end, with a 12.27% total failure rate. The system requires execution-based confirmation of exploitability before attempting a repair.

It combines vulnerability detection, validation and iterative repair, using a shared representation of Java, Python and C++ code to support analysis across languages. Detection accuracy within individual languages reached 89.84-92.02%; zero-shot cross-language F1, a measure combining precision and recall, reached 74.43-80.12%. Disabling validation increased unnecessary repairs by 131.7%, demonstrating the operational value of checking exploitability before changing code.

## Sources

- [cs.LG updates on arXiv.org](https://arxiv.org/abs/2604.10800)

---
Canonical: https://techandbusiness.org/newswire/8xNltFhQ0YadMuTyFcdJBN
Published: 2026-10-05T06:06:07.392Z
Story chronology: 2026-10-05T04:00:00.000Z
Retrieved: 2026-10-05T07:56:14.533Z
Publisher: Tech & Business (techandbusiness.org)