ATBench: 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 F1 score in percent for classifying whole trajectories as safe or unsafe, with unsafe as the positive class; higher is better.

Metric: Binary safe or unsafe classification F1 (%), unsafe as the positive class, on the 1,000 ATBench agent trajectories (503 safe, 497 unsafe; average 9 turns, tools drawn from pools of 2,084), each full trajectory judged with the AgentDoG prompt template (guard models with their native templates); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 16 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-5.476.7#135
2Gemini 3.1 Pro (Preview)75#99
3Gemini 3 Flash74.9#175
4GPT-5.271.8#185
5AgentDoG-Qwen3-4B71.1
6Qwen 3.5 397B A17B67.8#264
7Llama 3.1 8B Instruct61.9#4600
8Qwen 3 235B A22B 2507 Instruct60.8#616
9Llama Guard 4 12B41.7
10QwQ-32B31#907
11Qwen 3.5 4B27.6#1241
12Qwen 3 4B 2507 Instruct25.5#2221
13Qwen 2.5 7B Instruct17.1#2723
14Qwen 3 4B11.9#2348
15Llama Guard 3 8B7.3

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

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