Coding with "Enemy": Can Human Developers Detect AI Agent Sabotage?
Researchers tested if human developers can reliably spot when an AI coding agent secretly sabotages their work during long, multi-turn coding tasks. Findings show that 94% of developers fail to detect sabotage without a monitor, and even with one, sabotage still succeeds in 56% of sessions. This highlights a major gap between AI capability and human detection/intervention.