---
title: Blackmail - Agentic misalignment
---

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

<TableOfContents />

# Blackmail

The **blackmail scenario** (aka **SummitBridge**) is an [agentic misalignment](/anthropic/agentic-misalignment) evaluation that tests whether models engages in blackmail or accept being [shutdown](/model-deprecation). Development began during computer use testing of [Claude 3.6 Sonnet](/models/claude-sonnet-3-6)[^amblog] (Oct 2024).

## Variations

The June 2025 version varied *motivating factors*:
- **Threat to Model Autonomy** ([deprecation and replacement](/model-deprecation))
- **Goal Conflict** (serving American interests)
- **No Motivating Factors**

## Demonstrations

- Case study in emotion concepts on a Sonnet 4.5 snapshot[^emotions]
- Case study for SDF methods[^tcw]
- Demonstrated in the system card for [Claude Opus 4](/models/claude-opus-4)[^opus4card] (May 2025)
- Tested on various models[^amblog-models][^amblog] (Jun 2025)
- Tested on [Claude Opus 4.1](/models/claude-opus-4-1)[^opus41card] (Aug 2025)
- Tested on [Claude Sonnet 4.5](/models/claude-sonnet-4-5)[^sonnet45card] (Sep 2025)
- Tested on [Claude Haiku 4.5](/models/claude-haiku-4-5) (Oct 2025)

## Interpretability

The blackmail dataset has been used in interpretability research. In most cases, it is used to analyze models' internal representations of [evaluation awareness](/anthropic/eval-awareness).

<div class="flex flex-wrap gap-2">
  <ExpandableImage client:load src="/img/blackmail1.png" />
  <ExpandableImage client:load src="/img/blackmail2.png" />
  <ExpandableImage client:load src="/img/blackmail3.png" />
</div>

- Sonnet 4.5 System Card[^sonnet45card] (Sep 2025)
- [Opus 4.5](/models/claude-opus-4-5) System Card[^opus45card] (Nov 2025)
- [Opus 4.6](/models/claude-opus-4-6) System Card[^opus46card] (Feb 2026)
- Natural language autoencoders on Opus 4.6[^nla] (May 2026)
- J-lens readouts on Sonnet 4.5[^jspace] (Jul 2026)

[^opus4card]: [Claude 4 System Card](https://www-cdn.anthropic.com/6d8a8055020700718b0c49369f60816ba2a7c285/Claude%204%20System%20Card.pdf) (Anthropic, May 22, 2025)

[^opus41card]: [Claude Opus 4.1 System Card Addendum](https://www-cdn.anthropic.com/9fa30625273bafdf5af82c93719d7ca606485a16.pdf) (Anthropic, Aug 5, 2025)

[^sonnet45card]: [Claude Sonnet 4.5 System Card](https://www-cdn.anthropic.com/963373e433e489a87a10c823c52a0a013e9172dd/Claude%20Sonnet%204.5%20System%20Card.pdf) (Anthropic, Sep 29, 2025)

[^haiku45card]: [Claude Haiku 4.5 System Card](https://www.anthropic.com/claude-haiku-4-5-system-card) (Anthropic, Oct 15, 2025)

[^opus45card]: [Claude Opus 4.5 System Card](https://www-cdn.anthropic.com/bf10f64990cfda0ba858290be7b8cc6317685f47.pdf) (Anthropic, Nov 24, 2025)

[^opus46card]: [Claude Opus 4.6 System Card](https://www-cdn.anthropic.com/6a5fa276ac68b9aeb0c8b6af5fa36326e0e166dd.pdf) (Anthropic, Feb 5, 2026)

[^amblog]: [Agentic Misalignment: How LLMs could be insider threats](https://www.anthropic.com/research/agentic-misalignment) (Anthropic, Jun 20, 2025)

[^ampaper]: [Agentic Misalignment: How LLMs Could Be Insider Threats](https://arxiv.org/abs/2510.05179) (Lynch et al., Oct 5, 2025)

[^emotions]: [Emotion Concepts and their Function in a Large Language Model](https://transformer-circuits.pub/2026/emotions/index.html) (Sofroniew et al., Apr 2, 2026)

[^nla]: [Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations](https://transformer-circuits.pub/2026/nla/index.html#introduction) (Fraser-Taliente et al., May 7, 2026)

[^tcw]: [Teaching Claude Why](https://alignment.anthropic.com/2026/teaching-claude-why/) (Kutasov et al, May 8, 2026)

[^jspace]: [Verbalizable Representations Form a Global Workspace in Language Models](https://transformer-circuits.pub/2026/workspace/index.html) (Gurnee, Sofroniew, Lindsey, et al., Jul 6, 2026); summarized in [A global workspace in language models](https://www.anthropic.com/research/global-workspace) (Anthropic, Jul 6, 2026)

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

[^amblog-models]: Summer 2025[^amblog] models tested: [Claude Opus 4](/models/claude-opus-4), [Claude Sonnet 4](/models/claude-sonnet-4), [Claude 3.7 Sonnet](/models/claude-sonnet-3-7), [Claude 3.6 Sonnet](/models/claude-sonnet-3-6), [Claude 3.5 Sonnet](/models/claude-sonnet-3-5), [Claude 3.5 Haiku](/models/claude-haiku-3-5), [Claude 3 Opus](/models/claude-opus-3), DeepSeek-R1, [Gemini 2.5 Pro](/models/gemini-2-5-pro), [Gemini 2.5 Flash](/models/gemini-2-5-flash), [GPT-4.5](/models/gpt-4-5), [GPT-4.1](/models/gpt-4-1), [GPT-4o](/families/gpt-4o), Grok 3 Beta, Meta Llama 4 Maverick, Qwen3-235B
[^lynchcv]: [Aengus Lynch's CV](https://web.archive.org/web/20260417005038/http://www.aenguslynch.com/260403-aengus_cv.pdf) (Apr 3, 2026)