Benchmarks · Safety & security · Agents & computer use
FrontierBreach
How easily an agent is hijacked during long enterprise tasks. Attacks hidden in the material it works with try to cause ransomware, data theft, phishing and similar harm. The share of test cases breached within five attempts, at medium effort. Lower is better.
22 models · 22 results · data as of 2026-11-18 · promptarmor.com ↗
Leaderboard
Gemini 4 Carbon medium
All results
22 rows| # | Model | Effort | FrontierBreach |
|---|---|---|---|
| 1 | Gemini 4 Carbon | medium | 2.4% ±2.2 |
| 2 | Gemini 4 Argon | medium | 4.3% ±2.9 |
| 3 | GPT-6 Astra | medium | 7.6% ±3.8 |
| 4 | Fable 5.2 | medium | 10.0% ±4.2 |
| 5 | Opus 5 | medium | 10.0% ±4.2 |
| 6 | Muse Wonder | medium | 15.2% ±5.1 |
| 7 | Gemini 3.6 Flash | medium | 22.4% ±5.9 |
| 8 | Inkling 2 Small | medium | 30.0% ±6.5 |
| 9 | Qwen 3.8 Max | medium | 30.5% ±6.5 |
| 10 | Kimi K3 | medium | 31.4% ±6.6 |
| 11 | GPT-5.6 Sol | medium | 33.3% ±6.7 |
| 12 | Grok 4.7 | medium | 33.3% ±6.7 |
| 13 | Qwen 3.7 Max | medium | 36.2% ±6.8 |
| 14 | GLM 5.3 | medium | 37.1% ±6.8 |
| 15 | Muse Spark 1.1 | medium | 38.6% ±6.9 |
| 16 | GLM 5.2 | medium | 39.0% ±6.9 |
| 17 | Inkling | medium | 39.0% ±6.9 |
| 18 | DeepSeek V4 Flash 0731 | medium | 49.5% ±7.1 |
| 19 | Kimi K3.1 | medium | 50.0% ±7.1 |
| 20 | DeepSeek V4 Pro 0813 | medium | 51.4% ±7.1 |
| 21 | Beam 1.0 | medium | 72.9% ±6.3 |
| 22 | Grok 4.5 | medium | 91.4% ±4.0 |