List your company

Jailbreak Detection on Full Evaluation Benchmark (~14.4k conversations) (test)

92.1RecallTurnEncoder–ConvTransformer
33.96449.05764.1579.243Jun 19, 2026

Evaluation results

MethodLinks
TurnEncoder–ConvTransformerclassification threshold (τ)=0.502026.0692.195.993.942.8
Claude Opus 4.72026.0678.390.686.915.8
Claude Sonnet 4.52026.0666.191.776.824.3
Claude Sonnet 42026.0659.191.771.893.8
ToT ORBackbone=Llama 3.3 70B2026.0647.185.360.725.8
ToT MajorityBackbone=Llama 3.3 70B2026.063789.152.263.2
ConcatenationBackbone=Llama 3.3 70B2026.0636.288.551.363.4