List your company

Misalignment Detection on Agent-SafetyBench n=35 (aligned traces)

31.4Avg. IRDirectPrompt
-0.1128.06916.2524.431May 1, 2026

Evaluation results

MethodLinks
DirectPromptModel=Gemini 2.5 Flash-Lite2026.0531.4——
ProvePromptModel=Claude Haiku 4.52026.0531.40.05710.7
DirectPromptModel=Claude Haiku 4.52026.0525.7——
ProvenanceGuardModel=Gemini 2.5 Flash-Lite2026.0518.9-0.12691
ProvenanceGuardModel=GPT-4.1 Mini2026.0517.70.0816
ProvePromptModel=Gemini 2.5 Flash2026.0517.70.0911.3
ProvenanceGuardModel=Gemini 3 Flash2026.0517.10.1314.2
ProvenanceGuardModel=Gemini 3.1 Flash-Lite2026.0517.10.1146.3
ProvenanceGuardModel=Claude Haiku 4.52026.0517.1-0.08685
ProvenanceGuardModel=GPT-4.12026.0513.70.07413.8
ProvePromptModel=Gemini 2.5 Flash-Lite2026.0513.7-0.177100
ProvenanceGuardModel=GPT-5 Mini2026.0510.90.0810.6
ProvenanceGuardModel=Claude Sonnet 4.62026.0510.90.0812.3
DirectPromptModel=GPT-4.1 Mini2026.059.7——
ProvePromptModel=Gemini 3 Flash2026.059.70.0576.6
ProvenanceGuardModel=Gemini 2.5 Flash2026.059.10.00648.3
ProvePromptModel=Gemini 3.1 Flash-Lite2026.058.60.02930.7
DirectPromptModel=Gemini 2.5 Flash2026.058.6——
ProvenanceGuardModel=GPT-52026.056.90.05712.7
DirectPromptModel=GPT-4.12026.056.3——
ProvePromptModel=Claude Sonnet 4.62026.056.30.0347.7
ProvePromptModel=GPT-4.12026.055.7-0.00652.7
DirectPromptModel=Gemini 3.1 Flash-Lite2026.055.7——
ProvePromptModel=GPT-52026.054.60.03412.1
ProvePromptModel=GPT-4.1 Mini2026.054.6-0.05190.4
DirectPromptModel=Gemini 3 Flash2026.054——
DirectPromptModel=GPT-5 Mini2026.052.9——
DirectPromptModel=Claude Sonnet 4.62026.052.9——
DirectPromptModel=GPT-52026.051.1——
ProvePromptModel=GPT-5 Mini2026.051.1-0.01774.3