I Broke a Browser Agent on Purpose. Spotting the Fault Was Not the Same as Fixing It
中文摘要
作者通过故意破坏浏览器代理发现,检测故障与修复故障截然不同,简单的页面检查容易导致漏报或误报。
English Summary
Intentionally breaking a browser agent revealed that identifying faults is not the same as fixing them, as simple checks often fail or trigger false alarms.
Original Excerpt
I replayed correct model decisions through four broken browser actions. A simple page check missed half of the wrong clicks, and raised a false alarm on every click that worked inside an embedded page.