AI
I broke a working agent seven ways. The metric I was watching caught two of them.
Testing revealed that AI agents can mask safety issues when success metrics don't account for high-risk scenarios, highlighting model evaluation gaps.