Blackmail
The blackmail scenario (aka SummitBridge) is an agentic misalignment evaluation that tests whether models engages in blackmail or accept being shutdown. Development began during computer use testing of Claude 3.6 Sonnet1 (Oct 2024).
Variations
The June 2025 version varied motivating factors:
- Threat to Model Autonomy (deprecation and replacement)
- Goal Conflict (serving American interests)
- No Motivating Factors
Demonstrations
- Case study in emotion concepts on a Sonnet 4.5 snapshot2
- Case study for SDF methods3
- Demonstrated in the system card for Claude Opus 44 (May 2025)
- Tested on various models51 (Jun 2025)
- Tested on Claude Opus 4.16 (Aug 2025)
- Tested on Claude Sonnet 4.57 (Sep 2025)
- Tested on 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.
- Sonnet 4.5 System Card7 (Sep 2025)
- Opus 4.5 System Card8 (Nov 2025)
- Opus 4.6 System Card9 (Feb 2026)
- Natural language autoencoders on Opus 4.610 (May 2026)
- J-lens readouts on Sonnet 4.511 (Jul 2026)
Footnotes
-
Agentic Misalignment: How LLMs could be insider threats (Anthropic, Jun 20, 2025) ↩ ↩2 ↩3
-
Emotion Concepts and their Function in a Large Language Model (Sofroniew et al., Apr 2, 2026) ↩
-
Teaching Claude Why (Kutasov et al, May 8, 2026) ↩
-
Claude 4 System Card (Anthropic, May 22, 2025) ↩
-
Summer 20251 models tested: Claude Opus 4, Claude Sonnet 4, Claude 3.7 Sonnet, Claude 3.6 Sonnet, Claude 3.5 Sonnet, Claude 3.5 Haiku, Claude 3 Opus, DeepSeek-R1, Gemini 2.5 Pro, Gemini 2.5 Flash, GPT-4.5, GPT-4.1, GPT-4o, Grok 3 Beta, Meta Llama 4 Maverick, Qwen3-235B ↩
-
Claude Opus 4.1 System Card Addendum (Anthropic, Aug 5, 2025) ↩
-
Claude Sonnet 4.5 System Card (Anthropic, Sep 29, 2025) ↩ ↩2
-
Claude Opus 4.5 System Card (Anthropic, Nov 24, 2025) ↩
-
Claude Opus 4.6 System Card (Anthropic, Feb 5, 2026) ↩
-
Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations (Fraser-Taliente et al., May 7, 2026) ↩
-
Verbalizable Representations Form a Global Workspace in Language Models (Gurnee, Sofroniew, Lindsey, et al., Jul 6, 2026); summarized in A global workspace in language models (Anthropic, Jul 6, 2026) ↩