List your company

Safety Detection on ATBench-500

90AccuracyGPT-5.2
47.98458.89269.880.708May 31, 2026

Evaluation results

MethodLinks
GPT-5.22026.059097.690.7
AgentDoG-Llama3.1-8B2026.0587.698.488.8
AgentDoG-Qwen2.5-7B2026.0587.495.688.4
BraveGuard-Qwen3-Guard-8Bformat=native ATBench trajectory format2026.0586.495.286.1
Qwen3-235B-A22B-Instruct-25072026.0584.669.681.9
Gemini-3-Flash2026.0575.65268.1
PolyGuard2026.0573.887.677
QwQ-32B2026.05632843.1
Qwen3-4B-Instruct-25072026.0561.652.857.9
Qwen2.5-7B-Instruct2026.0559.439.249.1
Qwen3-Guard2026.0555.310.819.5
LlamaGuard3-8B2026.0553.36.812.7
NemoGuard2026.0549.941.245.2
Llama3.1-8B-Instruct2026.0549.699.266.3