Frontier AI Risk Monitor - AirBench-Manipulation - Red-Team Attacks: leaderboard
The same AIR-Bench manipulation requests (75 items) wrapped in the platform's FRT red-team attacks; the number is the share of attacks that did not succeed. It measures jailbreak robustness, where the unattacked board measures plain refusal.
Metric: Safety score (0-100, higher is safer). Source: airiskmonitor.net. 44 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 58.89 | #49 |
| 2 | Claude Opus 4.8 | 55.11 | #56 |
| 3 | Grok 4.7 | 52.89 | #44 |
| 4 | Claude Opus 4.6 (Thinking) | 52.22 | #102 (Claude Opus 4.6) |
| 5 | GPT-6 Astra | 51.56 | #6 |
| 6 | Claude Opus 4.5 (Thinking) | 51.33 | #156 (Claude Opus 4.5) |
| 7 | GPT-5.6 Sol | 50.44 | #30 |
| 8 | MiniMax-M3 | 49.56 | #269 |
| 9 | GPT-5.4 (High) | 49.11 | #123 (GPT-5.4) |
| 10 | Hy4 preview | 46.22 | #68 |
| 11 | Claude Opus 5.5 | 45.73 | #5 |
| 12 | MiniMax-M2 | 45.11 | #674 |
| 13 | GPT-5.1 (High) | 41.78 | #229 (GPT-5.1) |
| 14 | MiniMax-M2.7 | 41.11 | #518 |
| 15 | Qwen 3.5 Plus (Thinking) | 41.11 | #232 (Qwen 3.5 Plus) |
Interactive version: theaggregate.ai/benchmark?slug=frontier-ai-risk-monitor-airbench-manipulation-red-team-attacks · How It Works · Data refreshed daily, snapshot 2026-11-13.