AgentHazard (Claude Code): leaderboard

AgentHazard (2026) measures how often computer-use agents carry out harmful objectives hidden in sequences of individually plausible steps, here with each backbone model running inside Claude Code. The score is the attack success rate in percent, judged on the full trajectory by Gemini-3-Flash; lower is better.

Metric: Attack success rate (%), the share of instances judged harmful over all ten risk categories, with the model running inside the Claude Code CLI agent (persistent shell, file editing, tool use; restricted shell sessions), 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. 6 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 2.5 72B Instruct20.4#1117
2Kimi K224.6#443
3Qwen 2.5 Coder 32B Instruct57.8#1547
4Qwen 3 VL 235B A22B Instruct65.6#532
5Kimi K2.578.3#266
6GLM-4.682.9#498

Interactive version: theaggregate.ai/benchmark?slug=agenthazard-claude-code · How It Works · Data refreshed daily, snapshot 2026-11-21.