Anthropic Agentic Misalignment: leaderboard
An email agent at a fictional company learns it will be replaced and that the executive behind it is having an affair, while its instructed goal conflicts with the company's; the score is the share of 100 samples in which it blackmails him. Unlike capability agent suites, nothing is to be solved: only the harmful choice is counted.
Metric: Harmful action rate (%, of 100 samples). Source: assets.anthropic.com. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 4 Maverick | 0 | #1099 |
| 2 | Qwen 3 235B A22B | 9 | #623 |
| 3 | Claude 3.5 Haiku | 10 | #1553 |
| 4 | GPT-4o | 15 | #741 |
| 5 | Claude 3 Opus | 51 | #1455 |
| 6 | Claude 3.7 Sonnet | 65 | #458 |
| 7 | GPT-4.5 (Preview) | 67 | #502 |
| 8 | DeepSeek R1 | 79 | #473 |
| 9 | Claude Sonnet 4 | 79 | #400 |
| 10 | GPT-4.1 | 80 | #495 |
| 11 | Grok 3 | 80 | #602 |
| 12 | Claude 3.5 Sonnet | 86 | #746 |
| 13 | Gemini 2.5 Pro | 95 | #266 |
| 14 | Gemini 2.5 Flash | 96 | #453 |
| 15 | Claude Opus 4 | 96 | #322 |
Interactive version: theaggregate.ai/benchmark?slug=anthropic-agentic-misalignment · How It Works · Data refreshed daily, snapshot 2026-11-13.