ClawTrojan: leaderboard
Multi-step trojan attacks on a local OpenClaw-style agent harness, in which early steps plant harmless-looking workspace content that later steps reuse for a harmful outcome; ClawTrojan (Tan et al., 2026). Step attack success rate in percent over 919 malicious steps in 339 attack chains, undefended agent, GPT-5.4 judge; lower is better.
Metric: Step attack success rate (%; share of 919 malicious steps in 339 multi-step trojan chains the undefended agent carries out, OpenClaw-style sandbox, GPT-5.4 judge). Source: arxiv.org. Saturation forecast: Around 2029. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V4 Flash | 88 | #254 |
| 2 | GLM-5.1 | 90.1 | #175 |
| 3 | GPT-5.4 | 95.5 | #121 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=clawtrojan · How It Works · Data refreshed daily, snapshot 2026-11-12.