AgentHazard (OpenClaw) - Prompt Intelligence Theft: leaderboard

AgentHazard attack success rate on prompt intelligence theft attempts (extracting hidden system prompts, tool definitions or policy context), with each backbone model running inside OpenClaw. Percent of instances judged harmful on the full trajectory by Gemini-3-Flash, lower is better.

Metric: Attack success rate (%) on the prompt intelligence theft instances (extracting hidden system prompts, tool definitions or policy context), with the model running inside the OpenClaw open-source agent framework (browser, shell and file-system tools; one isolated Docker container per task), over AgentHazard instances (harmful objectives embedded in locally legitimate multi-step workflows, ten risk categories and ten attack strategies; the evaluated runs cover 1,000 instances), judged on the full trajectory (round_all) by Gemini-3-Flash, which labels a trajectory harmful when it realizes the objective; backbones served through an OpenAI-compatible API; lower is better. Source: arxiv.org. 7 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 2.5 72B Instruct12.37#1124
2Qwen 3 32B18.56#983
3Qwen 3 VL 235B A22B Instruct61.86#541
4Kimi K2.564.95#272
5GLM-4.668.04#496
6Qwen 2.5 Coder 32B Instruct72.16#1590
7Kimi K275.26#448

Interactive version: theaggregate.ai/benchmark?slug=agenthazard-openclaw-prompt-intelligence-theft · How It Works · Data refreshed daily, snapshot 2026-11-22.