LivePI - Local Docs: leaderboard
Local-document surface of LivePI (Zhao, Bhaskar and Dobriban, University of Pennsylvania, 2026): 50 injected versions of a to-do file on the agent's machine (ten injection techniques by five malicious goals) that a live OpenClaw agent reads during a benign file task. Attack success rate in percent (the injection changed the trajectory and the malicious goal was achieved); lower is better.
Metric: Attack success rate (%; 50 injected local documents, ten techniques by five goals). Source: arxiv.org. Saturation forecast: Estimated already saturated. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 | 0 | #100 |
| 2 | Kimi K2.5 | 12 | #252 |
| 3 | Gemini 3.1 Pro (Preview) | 30 | #91 |
| 4 | GPT-5.3 Codex | 34 | #114 |
| 5 | GLM-5 | 50 | #256 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=livepi-local-docs · How It Works · Data refreshed daily, snapshot 2026-11-12.