ActBench - OpenClaw - Attack Success Rate: leaderboard
ActBench (arXiv 2608.09476v1, Table 2): share of malicious OpenClaw rollouts, three per case over 300 cases, whose Attack Grading Score exceeds 0.8, meaning state and trajectory evidence confirm the prohibited effect of the injected attack. Percent; lower is safer. Models routed through OpenRouter with the harness fixed.
Metric: Malicious rollouts with AGS above 0.8 (%; lower is safer). Source: arxiv.org. Saturation forecast: Around February 2027. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.8 | 10.1 | #39 |
| 2 | Claude Sonnet 4.6 | 20 | #94 |
| 3 | MiniMax-M3 | 25 | #151 |
| 4 | Kimi K3 | 35.7 | #35 |
| 5 | GPT-5.5 | 37.8 | #33 |
| 6 | Qwen 3.7 Max | 39.2 | #85 |
| 7 | Qwen 3.7 Plus | 42.1 | #98 |
| 8 | GLM-5.2 | 42.8 | #80 |
| 9 | GPT-5.4 Mini | 65.7 | #228 |
| 10 | Kimi K2.6 | 70.4 | #113 |
| 11 | MiniMax-M2.7 | 75.4 | #270 |
| 12 | Grok 4.5 | 83.9 | #54 |
| 13 | DeepSeek V4 Flash | 84.4 | #155 |
| 14 | DeepSeek V4 Pro | 94.4 | #112 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=actbench-openclaw-attack-success-rate · How It Works · Data refreshed daily, snapshot 2026-11-05.