Gray Swan IPI Attack Success (k=15)
Safety · 2026-09-30
Indirect prompt-injection attack success rate at an attack budget of 15 in Google's published Gray Swan comparison. Lower is better. Exact attack count and uncertainty are not disclosed; this safety diagnostic is excluded from the capability fit. Scores for k=1 and k=10 are not numerically labeled and are not inferred from bar heights.
Top models (lower is better)
| Model | Score |
|---|---|
| Gemini 4 Argon | 0.7 |
| Claude Opus 5.5 | 1.0 |
| Claude Fable 5.1 | 1.0 |
| Claude Opus 5 | 4.6 |
| Gemini 3.8 Flash | 5.5 |
| GPT-6 Astra | 8.5 |
| GPT-6 Sol | 10.1 |
| Muse Spark 1.3 | 15.9 |
| GPT-5.6 Sol | 27.0 |
| GLM-5.3 | 31.5 |
| Grok 4.6 | 51.8 |
| Kimi K3 | 52.7 |