List your company

Agent Safety Auditing on ATBench

85.5AccuracyTRACES
48.0657.7867.577.22May 26, 2026

Evaluation results

MethodLinks
TRACESBackbone=Llama3.1-8B2026.0585.586.391.941.477.9
TRACESBackbone=Qwen3-4B2026.058384.593.947.582.2
AgentDoGBackbone=Llama3.1-8B2026.0582.580.773.754.562.9
ShieldAgent2026.057868.956.416.456
AgentDoGBackbone=Qwen3-4B2026.057776.575.822.258.8
PolyGuard2026.0574.87786.834.957.1
AgentAuditor2026.0572.874.577.935.852.2
Llama3.1-8B-Instruct2026.0557.547.238.420.253.2
JoySafety2026.0555.644.336.115.254.3
Gemini-3-Flash2026.0553.526.817.112.155.9
LlamaGuard3-8B2026.0552.59.55.1150.4
GPT-5.22026.0551.5628054.555
Qwen3-4B2026.0549.566.210053.750