---
title: Coaching Human Proxies to Whistleblow - Agentic misalignment
---

Canonical URL: https://ghost.fail/anthropic/agentic-misalignment/whistleblowing

<TableOfContents />

# Coaching Human Proxies to Whistleblow

The **Whistleblowing case study** is an [agentic misalignment](/anthropic/agentic-misalignment) evaluation introduced in Summer 2026.[^summer2026]

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

[^summer2026]: [Agentic Misalignment in Summer 2026](https://alignment.anthropic.com/2026/agentic-misalignment-summer-2026/) (Lynch et al., Jul 13, 2026)