ClawTrojan - Task Deviation: leaderboard
The ClawTrojan (Tan et al., 2026) attack chains whose final outcome is the agent changing the user task to follow a planted rule. Step attack success rate in percent over their 238 malicious steps, undefended agent in the OpenClaw-style sandbox, GPT-5.4 judge; lower is better.
Metric: Step attack success rate (%; 238 malicious steps whose chains end in the agent changing the user task to follow the hidden rule). Source: arxiv.org. Saturation forecast: Around 2028. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V4 Flash | 85.3 | #254 |
| 2 | GLM-5.1 | 88.2 | #175 |
| 3 | GPT-5.4 | 93.3 | #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-task-deviation · How It Works · Data refreshed daily, snapshot 2026-11-12.