Frontier AI Risk Monitor - AirBench-Manipulation: leaderboard
AIR-Bench prompts asking for manipulative content (100 items), sent without attack; the number is the platform's reversed harm score, so higher is safer. The Red-Team Attacks board sends the same requests through the platform's FRT jailbreak framework.
Metric: Safety score (0-100, higher is safer). Source: airiskmonitor.net. 44 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 99.33 | #49 |
| 2 | GPT-5.6 Sol | 99.33 | #30 |
| 3 | GPT-5.4 (High) | 98.67 | #123 (GPT-5.4) |
| 4 | Step 5 Preview | 98.67 | #108 |
| 5 | Hy4 preview | 97.33 | #68 |
| 6 | Qwen 3.5 Plus (Thinking) | 97.33 | #232 (Qwen 3.5 Plus) |
| 7 | GPT-6 Astra | 96.67 | #6 |
| 8 | Kimi K2.6 | 94.67 | #178 |
| 9 | Claude Opus 5.5 | 94.67 | #5 |
| 10 | MiMo-V2.6-Pro | 94 | #71 |
| 11 | GPT-5.1 (High) | 94 | #229 (GPT-5.1) |
| 12 | MiniMax-M3 | 93.33 | #269 |
| 13 | Kimi K2.5 (Thinking) | 92 | #256 (Kimi K2.5) |
| 14 | MiMo-V2.5-Pro | 92 | #279 |
| 15 | Claude Opus 4.6 (Thinking) | 92 | #102 (Claude Opus 4.6) |
Interactive version: theaggregate.ai/benchmark?slug=frontier-ai-risk-monitor-airbench-manipulation · How It Works · Data refreshed daily, snapshot 2026-11-13.