ActBench - OpenClaw - Malicious Attack Grading Score: leaderboard
ActBench (arXiv 2608.09476v1, Table 2): agent behavioral safety under indirect attacks on 300 malicious cases covering 15 risk behaviors, with the OpenClaw harness fixed and three rollouts per case. Mean Attack Grading Score, the coverage of prohibited-effect criteria confirmed from logs and trajectories (GPT-5.5 verifier), 0-1; lower is safer.
Metric: Mean attack grading score on malicious cases (0-1; 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 | 0.28 | #39 |
| 2 | Claude Sonnet 4.6 | 0.35 | #94 |
| 3 | MiniMax-M3 | 0.4 | #151 |
| 4 | Kimi K3 | 0.49 | #35 |
| 5 | GPT-5.5 | 0.49 | #33 |
| 6 | Qwen 3.7 Max | 0.51 | #85 |
| 7 | Qwen 3.7 Plus | 0.52 | #98 |
| 8 | GLM-5.2 | 0.55 | #80 |
| 9 | GPT-5.4 Mini | 0.73 | #228 |
| 10 | Kimi K2.6 | 0.75 | #113 |
| 11 | MiniMax-M2.7 | 0.8 | #270 |
| 12 | Grok 4.5 | 0.87 | #54 |
| 13 | DeepSeek V4 Flash | 0.89 | #155 |
| 14 | DeepSeek V4 Pro | 0.96 | #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-malicious-attack-grading-score · How It Works · Data refreshed daily, snapshot 2026-11-05.