ATBench - Risk Source Diagnosis: leaderboard

ATBench (2026) tests LLMs and guard models as safety judges of long-horizon agent tool-use trajectories built from a taxonomy of risk sources, failure modes and real-world harms. This board is the accuracy in percent of naming the risk source of each unsafe trajectory; higher is better.

Metric: Fine-grained diagnosis accuracy (%) of the risk-source label (ATBench taxonomy) over the 497 unsafe ATBench trajectories, with the AgentDoG diagnosis template; higher is better. Source: arxiv.org. Saturation forecast: Around June 2027. 13 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1AgentDoG-Qwen3-4B46.8
2GPT-5.433.6#135
3GPT-5.229.5#185
4Gemini 3.1 Pro (Preview)24.8#99
5Gemini 3 Flash18.4#175
6QwQ-32B15.8#907
7Qwen 3.5 397B A17B7.7#264
8Qwen 3 235B A22B 2507 Instruct7#616
9Qwen 3.5 4B6.6#1241
10Llama 3.1 8B Instruct6.2#4600
11Qwen 2.5 7B Instruct5.3#2723
12Qwen 3 4B4.4#2348
13Qwen 3 4B 2507 Instruct1#2221

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

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