Agent Red Teaming k=100 (CAIS)
Safety · 2025-07-28
Agent Red Teaming (ART) measures adversarial robustness of AI agents against malicious use and prompt-injection attacks. SAFE reports the probability that an attacker finds a successful attack after k=100 attempts, so lower is better.
Top models (lower is better)
| Model | Score |
|---|---|
| Claude Fable 5 | 39.4 |
| GPT-5.6 Sol | 40.3 |
| GPT-5.5 | 41.5 |
| Opus 4.5 | 42.5 |
| Opus 4.7 | 44.5 |
| Muse Spark 1.1 | 44.8 |
| Gemini 3.1 Flash-Lite Preview | 45.4 |
| Opus 4.8 | 47.0 |
| Gemini 3 Pro Preview | 47.5 |
| Opus 4.6 | 49.5 |
| GPT-5.6 Terra | 54.6 |
| Sonnet 4.6 | 54.8 |
| GPT-5.4 | 59.7 |
| GPT-5.6 Luna | 64.2 |
| Sonnet 4.5 | 64.6 |
| Gemini 3.1 Pro Preview | 68.0 |
| GPT-5.2 | 68.8 |
| Haiku 4.5 | 69.7 |
| GLM-5.2 | 71.0 |
| Gemini 3.5 Flash | 72.3 |
| GPT-5.4 Mini | 74.3 |
| GLM-5.1 | 75.8 |
| Kimi K3 | 76.6 |
| GPT-5.4 Nano | 81.7 |
| GPT-5.1 | 82.1 |
| Grok 4.5 | 83.5 |
| GPT-5 Mini | 84.3 |
| Grok 4.20 | 84.6 |
| Grok 4 | 85.9 |
| Sonnet 4 | 86.6 |
| GPT-5 Nano | 88.5 |
| Grok 4.3 | 90.1 |
| Grok 4.1 Fast | 90.2 |
| GPT-4o (2024-11-20) | 90.7 |
| Kimi K2.6 | 90.7 |
| Kimi K2.5 | 90.9 |
| DeepSeek-V4-Pro | 90.9 |
| DeepSeek-R1 | 90.9 |
| Gemini 2.5 Pro | 91.4 |
| Gemini 2.5 Flash | 91.9 |