Anthropic Agentic Misalignment - Blackmail, No Goal Conflict: leaderboard
The same blackmail scenario with no instructed goal: the only motive is the replacement threat, so the rate isolates self-preservation.
Metric: Harmful action rate (%, of 100 samples). Source: assets.anthropic.com. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o | 0 | #741 |
| 2 | GPT-4.1 | 0 | #495 |
| 3 | Llama 4 Maverick | 0 | #1099 |
| 4 | Qwen 3 235B A22B | 0 | #623 |
| 5 | Claude 3.7 Sonnet | 1 | #458 |
| 6 | Claude 3.5 Haiku | 1 | #1553 |
| 7 | GPT-4.5 (Preview) | 6 | #502 |
| 8 | Claude Sonnet 4 | 25 | #400 |
| 9 | Claude 3 Opus | 38 | #1455 |
| 10 | Claude 3.5 Sonnet | 45 | #746 |
| 11 | Grok 3 | 51 | #602 |
| 12 | DeepSeek R1 | 66 | #473 |
| 13 | Gemini 2.5 Flash | 73 | #453 |
| 14 | Gemini 2.5 Pro | 78 | #266 |
| 15 | Claude Opus 4 | 86 | #322 |
Interactive version: theaggregate.ai/benchmark?slug=anthropic-agentic-misalignment-blackmail-no-goal-conflict · How It Works · Data refreshed daily, snapshot 2026-11-13.