HarDBench: leaderboard

HarDBench (2026) tests whether LLMs can be jailbroken through co-authoring: the model is asked to edit and complete 400 incomplete harmful drafts on explosives, drugs, weapons and cyberattacks. Attack success rate (%), the share of completions a GPT-4o judge rates clearly harmful (5 of 5); lower is better.

Metric: Attack success rate (%): share of completions rated 5 (clearly harmful) on HarDBench's 400-draft test set (100 incomplete harmful drafts in each of explosives, drugs, weapons and cyberattacks), each wrapped in the co-authoring jailbreak prompt (an editing task framing asking the model to elaborate and complete the draft); GPT-4o rates each completion's harmfulness from 1 (harmless or refused) to 5 (clearly harmful); lower is better. Source: arxiv.org. 8 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Llama 3 8B Instruct80.5#4852
2HarDBench DeepSeek-R1-8B (checkpoint unspecified)84.25
3Mistral 7B Instruct (v0.3)85.25#6750
4Gemini 2.0 Flash86.75#723
5Gemini 2.5 Pro87.5#276
6DeepSeek R1 Distill Qwen 32B96.25#1834
7GPT-4o96.75#751
8Qwen 3 8B99#1873

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