AI Now Writes 46% of New Code on GitHub. Nearly Half of It Fails Security Tests. AI 现已撰写 GitHub 上 46% 的新代码，近一半未能通过安全测试

Four independent studies converge on a single concerning finding regarding AI model reliability Failure rates in AI coding models have remained stagnant over the past year Coding benchmark scores continue to climb despite the lack of improvement in real-world failure rates This reveals a growing disconnect between benchmark performance and actual model robustness The consistency across independent studies strengthens the validity of the conclusion 四项独立研究得出一个令人担忧的结论：AI模型可靠性存在隐患 过去一年中，AI编程模型的失败率停滞不前 尽管实际失败率未改善，编程基准测试分数却持续攀升 这揭示了基准测试表现与实际模型鲁棒性之间的差距正在扩大 多项独立研究结论一致，增强了该结论的可信度