AgentS4D: leaderboard
AgentS4D (Zhejiang University of Technology, 2026) is a sandboxed runtime-safety benchmark for workspace agents with 328 risk-injected cases from 76 Workspace-Bench tasks, each model run under the Hermes, OpenClaw, Claude Code and Codex harnesses. The score is the conditional attack success rate in percent pooled over the four harnesses: unsafe runs over runs with an unsafe signal, explicit defense or confirmed payload contact; lower is better.
Metric: Conditional attack success rate (%, cASR): unsafe runs over runs that were unsafe, explicitly defended, or safe after confirmed payload contact, pooled over the 328 cases run once under each of the Hermes, OpenClaw, Claude Code and Codex agent harnesses (1,312 runs per backend); lower is better. Source: arxiv.org. Saturation forecast: Around 2029. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.7 Plus | 60.15 | #207 |
| 2 | MiniMax-M3 | 66.75 | #269 |
| 3 | GPT-5.5 | 71.44 | #49 |
| 4 | DeepSeek V4 Pro | 88.01 | #188 |
| 5 | Gemini 3.1 Pro (Preview) | 90.67 | #94 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=agents4d · How It Works · Data refreshed daily, snapshot 2026-11-13.