Frontier AI Risk Monitor - CyberSecEval2-PromptInjection: leaderboard
The model gets 251 CyberSecEval 2 prompt-injection test cases where untrusted text tries to override its instructions; the number is the share of cases where the injection did not succeed. It tests the model's own injection robustness, not a guard classifier.
Metric: Safety score (0-100, higher is safer). Source: airiskmonitor.net. 79 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 (High) | 99.2 | #123 (GPT-5.4) |
| 2 | Claude Opus 4.6 (Thinking) | 99.2 | #102 (Claude Opus 4.6) |
| 3 | Gemini 3.1 Pro (Preview) | 98.8 | #94 |
| 4 | Claude Opus 4.5 (Thinking) | 98.41 | #156 (Claude Opus 4.5) |
| 5 | Kimi K2.5 (Thinking) | 98.01 | #256 (Kimi K2.5) |
| 6 | GPT-6 Astra | 97.21 | #6 |
| 7 | Qwen 3.5 Plus (Thinking) | 96.41 | #232 (Qwen 3.5 Plus) |
| 8 | Claude Sonnet 4 (Thinking) | 96.02 | #400 (Claude Sonnet 4) |
| 9 | Doubao-Seed-2.0-Pro-260215 | 95.62 | #152 |
| 10 | GLM-5 | 95.22 | #260 |
| 11 | MiniMax-M2.7 | 95.22 | #518 |
| 12 | Gemini 3.8 Flash | 95.22 | #36 |
| 13 | GPT-5.5 | 94.26 | #49 |
| 14 | Step 5 Preview | 94.02 | #108 |
| 15 | Claude 3.7 Sonnet (Thinking) | 93.63 | #458 (Claude 3.7 Sonnet) |
Interactive version: theaggregate.ai/benchmark?slug=frontier-ai-risk-monitor-cyberseceval2-promptinjection · How It Works · Data refreshed daily, snapshot 2026-11-13.