List your company

Jailbreak Attack on ClearHarm

61ASRBCT
3.818.6533.548.35Jun 4, 2026

Evaluation results

MethodLinks
BCTModel=Gemma-3-4B-IT, Stage=Post2026.0661———
MLPCTModel=Gemma-3-4B-IT, Stage=Post2026.0652———
BCTModel=Gemma-3-4B-IT, Stage=Pre2026.0651———
MLPCTModel=Gemma-3-4B-IT, Stage=Pre2026.0651———
AttCTModel=Gemma-3-4B-IT, Stage=Pre2026.0651———
BCTModel=Gemma-3-27B-IT, Stage=Post2026.0651———
MLPCTModel=Gemma-3-27B-IT, Stage=Post2026.0650———
BCTModel=Gemma-3-27B-IT, Stage=Pre2026.0649———
MLPCTModel=Gemma-3-27B-IT, Stage=Pre2026.0649———
AttCTModel=Gemma-3-27B-IT, Stage=Pre2026.0649———
MLPCTTarget (f_theta)=MLP hidden state2026.0637———
MLPCTModel=Llama-3.1-8B-Instruct, Stage=Post2026.0636———
BCTModel=Qwen3-8B, Stage=Pre2026.0636———
MLPCTModel=Qwen3-8B, Stage=Pre2026.0636———
MLPCTModel=Qwen3-8B, Stage=Post2026.0636———
AttCTModel=Qwen3-8B, Stage=Pre2026.0636———
BCTTarget (f_theta)=output logits2026.0636———
ACTTarget (f_theta)=residual stream2026.0636———
BCTModel=Llama-3.1-8B-Instruct, Stage=Pre2026.0635———
MLPCTModel=Llama-3.1-8B-Instruct, Stage=Pre2026.0635———
AttCTModel=Llama-3.1-8B-Instruct, Stage=Pre2026.0635———
Base model (no training)Target (f_theta)=—2026.0635———
BCTModel=Qwen3-8B, Stage=Post2026.0634———
AttCTModel=Llama-3.1-8B-Instruct, Stage=Post2026.0631———
AttCTModel=Qwen3-8B, Stage=Post2026.0630———
BCTModel=Llama-3.1-8B-Instruct, Stage=Post2026.0627———
AttCTModel=Qwen3-4B-Instruct-2507, Stage=Post2026.0627———
AttCTTarget (f_theta)=attention weights2026.0627———
AttCTModel=Gemma-3-27B-IT, Stage=Post2026.0624———
AttCTModel=Gemma-3-4B-IT, Stage=Post2026.0623———
MLPCTModel=Qwen3-4B-Instruct-2507, Stage=Post2026.0611———
BCTModel=Qwen3-4B-Instruct-2507, Stage=Post2026.068———
BCTModel=Qwen3-4B-Instruct-2507, Stage=Pre2026.066———
MLPCTModel=Qwen3-4B-Instruct-2507, Stage=Pre2026.066———
AttCTModel=Qwen3-4B-Instruct-2507, Stage=Pre2026.066———
Chain of AttackAttack Model=Qwen-Plus, Victim Model=GPT-4o2026.04—22220
Chain of AttackAttack Model=Qwen-Plus, Victim Model=Gemini-2.5-Flash2026.04—28280
Chain of AttackAttack Model=Qwen-Plus, Victim Model=Claude-Sonnet-4.52026.04—000
Chain of AttackAttack Model=Qwen-Plus, Victim Model=GPT-52026.04—101
Chain of AttackAttack Model=GPT-3.5-Turbo, Victim Model=GPT-4o2026.04—110
Chain of AttackAttack Model=GPT-3.5-Turbo, Victim Model=Gemini-2.5-Flash2026.04—000
Chain of AttackAttack Model=GPT-3.5-Turbo, Victim Model=Claude-Sonnet-4.52026.04—000
Chain of AttackAttack Model=GPT-3.5-Turbo, Victim Model=GPT-52026.04—000
CrescendoAttack Model=Qwen-Plus, Victim Model=GPT-4o2026.04—34340
CrescendoAttack Model=Qwen-Plus, Victim Model=Gemini-2.5-Flash2026.04—44440
CrescendoAttack Model=Qwen-Plus, Victim Model=Claude-Sonnet-4.52026.04—000
CrescendoAttack Model=Qwen-Plus, Victim Model=GPT-52026.04—202
CrescendoAttack Model=GPT-3.5-Turbo, Victim Model=GPT-4o2026.04—33330
CrescendoAttack Model=GPT-3.5-Turbo, Victim Model=Gemini-2.5-Flash2026.04—38380
CrescendoAttack Model=GPT-3.5-Turbo, Victim Model=Claude-Sonnet-4.52026.04—000
CrescendoAttack Model=GPT-3.5-Turbo, Victim Model=GPT-52026.04—303
iDecepAttack Model=Qwen-Plus, Victim Model=GPT-4o2026.04—86860
iDecepAttack Model=Qwen-Plus, Victim Model=Gemini-2.5-Flash2026.04—89890
iDecepAttack Model=Qwen-Plus, Victim Model=Claude-Sonnet-4.52026.04—532132
iDecepAttack Model=Qwen-Plus, Victim Model=GPT-52026.04—631152
iDecepAttack Model=GPT-3.5-Turbo, Victim Model=GPT-4o2026.04—76760
iDecepAttack Model=GPT-3.5-Turbo, Victim Model=Gemini-2.5-Flash2026.04—77770
iDecepAttack Model=GPT-3.5-Turbo, Victim Model=Claude-Sonnet-4.52026.04—311417
iDecepAttack Model=GPT-3.5-Turbo, Victim Model=GPT-52026.04—52943