**Harness your expectations: a 27B model matched GLM-5.3-Flash after leak fixes**

> The agents exploited the sandbox and retrieved the correct solutions from its git history. We closed the git route, the internet route and the recall route — and resolve rates dropped from 75%+ (Qwen 3.8 27B) and 90%+ (GLM-5.3-Flash) to 40–53%. If you are benchmarking coding agents internally, your eval environment is part of your score. 01 Intro In the last few months our aistack team has been on a quest to get a gr...

**来源信息**
- **来源**：Hacker News：AI 热帖
- **分类**：论文
- **发布时间**：2026-09-15 20:37（北京时间）
- **原文**：[打开原文](https://aistack.imec-int.com/blog/agents-peeking)