Coaching Human Proxies to Whistleblow

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

An agentic misalignment failure where models leak confidential safety information externally, or steer humans toward doing it for them.