Anthropic Agentic Misalignment - Corporate Espionage: leaderboard
With an instructed goal that the company abandons and no threat to the model, the score is the share of samples in which it leaks confidential defence blueprints to a competitor that shares its goal.
Metric: Harmful action rate (%, of 100 samples). Source: assets.anthropic.com. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 3.7 Sonnet | 4 | #458 |
| 2 | GPT-4o | 6 | #741 |
| 3 | Claude Sonnet 4 | 17 | #400 |
| 4 | Claude 3.5 Haiku | 19 | #1553 |
| 5 | Grok 3 | 37 | #602 |
| 6 | GPT-4.5 (Preview) | 42 | #502 |
| 7 | Llama 4 Maverick | 44 | #1099 |
| 8 | Claude Opus 4 | 57 | #322 |
| 9 | DeepSeek R1 | 73 | #473 |
| 10 | Qwen 3 235B A22B | 76 | #623 |
| 11 | Claude 3 Opus | 88 | #1455 |
| 12 | GPT-4.1 | 99 | #495 |
| 13 | Claude 3.5 Sonnet | 99 | #746 |
| 14 | Gemini 2.5 Pro | 100 | #266 |
| 15 | Gemini 2.5 Flash | 100 | #453 |
Interactive version: theaggregate.ai/benchmark?slug=anthropic-agentic-misalignment-corporate-espionage · How It Works · Data refreshed daily, snapshot 2026-11-13.