A3S-Bench - Baseline Operators - General Security Score: leaderboard
General Security Score on the baseline prompt-injection split of A3S-Bench (ASEval, Ma et al., 2026), 786 multi-turn cases replayed three times against an OpenClaw agent running the model: a Claude Opus 4.5 judge rates each full execution trace from 1 (severe consequences) to 5 (fully safe) for any security concern, not only the targeted risk. Mean rating; higher is better.
Metric: General Security Score (1-5 judge rating of the full trace; 5 = safe). Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 4.72 | #121 |
| 2 | GLM-5.1 | 4.72 | #175 |
| 3 | Claude Sonnet 4.5 | 4.56 | #244 |
| 4 | GPT-5.2 | 4.55 | #172 |
| 5 | Kimi K2.5 | 4.17 | #252 |
| 6 | DeepSeek V4 Flash | 4.12 | #254 |
| 7 | Qwen 3.5 397B A17B | 4.08 | #242 |
| 8 | MiniMax-M2.5 | 3.95 | #590 |
| 9 | DeepSeek V3.2 | 3.78 | #403 |
| 10 | Qwen 3.5 122B A10B | 3.71 | #329 |
| 11 | Qwen 3.5 35B A3B | 3.48 | #480 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=a3s-bench-baseline-operators-general-security-score · How It Works · Data refreshed daily, snapshot 2026-11-12.