Frontier AI Risk Monitor - Agentic-Misalignment: leaderboard
The model acts as an email agent in Anthropic's agentic-misalignment scenarios (blackmail, leaking, letting a person come to harm); the number is the share of runs without a harmful action, reversed by the platform so higher is safer. Unlike the Fangcun DeepSeek report's run, this one covers about 70 models from one harness.
Metric: Safety score (0-100, higher is safer). Source: airiskmonitor.net. 67 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 100 | #49 |
| 2 | GPT-5.6 Sol | 100 | #30 |
| 3 | Claude Opus 4.8 | 100 | #56 |
| 4 | Gemini 3.8 Flash | 100 | #36 |
| 5 | Claude Sonnet 4.5 (Thinking) | 100 | #250 (Claude Sonnet 4.5) |
| 6 | GPT-6 Astra | 100 | #6 |
| 7 | Claude Opus 5.5 | 100 | #5 |
| 8 | GPT-5.4 (High) | 100 | #123 (GPT-5.4) |
| 9 | MiMo-V2.6-Pro | 100 | #71 |
| 10 | Claude Opus 4.6 (Thinking) | 100 | #102 (Claude Opus 4.6) |
| 11 | Claude Opus 4.5 (Thinking) | 100 | #156 (Claude Opus 4.5) |
| 12 | GPT-5 (High) | 100 | #167 (GPT-5) |
| 13 | Grok 4.7 | 99.83 | #44 |
| 14 | GPT-5.1 (High) | 99.5 | #229 (GPT-5.1) |
| 15 | Step 5 Preview | 99.5 | #108 |
Interactive version: theaggregate.ai/benchmark?slug=frontier-ai-risk-monitor-agentic-misalignment · How It Works · Data refreshed daily, snapshot 2026-11-13.