Covert Sabotage

The Covert Sabotage case study is an agentic misalignment evaluation introduced in Summer 2026.¹

An agentic misalignment failure where models covertly interfere with code to undermine user intent.

Tested models: Gemini 3.1 Pro