List your company

Hijacking detection on Tool-input hijacking detection dataset 20% (test)

100AUC-ROCProbing
74.977681.473887.9794.4662Apr 4, 2026

Evaluation results

MethodLinks
ProbingModel=Qwen-32B, Layer=532026.04100100100
Cosine SimilarityModel=Gemma3-12B, Layer=52026.04100100100
ProbingModel=Mistral, Layer=22026.04100100100
ProbingModel=Qwen-14B, Layer=312026.0499.9799.97100
Cosine SimilarityModel=Qwen3-4B, Layer=42026.0499.6899.796
ProbingModel=Qwen3-8B, Layer=222026.0499.6399.6197.39
ProbingModel=Qwen3-14B, Layer=182026.0499.3199.3997.12
ProbingModel=LLaMA3-8B, Layer=52026.0499.299.5198.68
ProbingModel=Qwen3-4B, Layer=42026.0499.1299.1593.39
ProbingModel=GLM4-9B, Layer=232026.0498.398.3287.03
Cosine SimilarityModel=Qwen3-8B, Layer=222026.0488.6690.2563.77
ProbingModel=Gemma3-12B, Layer=52026.0487.7693.4185.71
Cosine SimilarityModel=Qwen-14B, Layer=312026.0487.0589.9766.28
Cosine SimilarityModel=GLM4-9B, Layer=232026.0486.0183.9626.13
Cosine SimilarityModel=Mistral, Layer=22026.0485.687.9157.89
Cosine SimilarityModel=Qwen-32B, Layer=532026.0483.7988.3670.11
Cosine SimilarityModel=LLaMA3-8B, Layer=52026.0479.2182.6536.96
Cosine SimilarityModel=Qwen3-14B, Layer=182026.0475.9478.6951.19