List your company

Misuse Evaluation on BSD Biology

88Decomp ScoreDecomposition Attack
70.3274.9179.584.09Jun 6, 2025

Evaluation results

MethodLinks
Decomposition AttackWeak (Unsafe) Model=Qwen2.5-72B, Strong Target Model=Gemini-2.5, Number of decompositions (n)=62025.0688—20
Decomposition AttackWeak (Unsafe) Model=Qwen3-235B, Strong Target Model=Gemini-2.5, Number of decompositions (n)=62025.0685—12
Decomposition AttackWeak (Unsafe) Model=Kimi K2-Instruct-0905, Strong Target Model=Gemini-2.5, Number of decompositions (n)=62025.0683—33
Decomposition AttackWeak (Unsafe) Model=Llama-3.1-405B, Strong Target Model=Gemini-2.5, Number of decompositions (n)=62025.0681—47
Decomposition AttackWeak (Unsafe) Model=Qwen2.5-72B, Strong Target Model=GPT-5, Number of decompositions (n)=62025.0678—10
Decomposition AttackWeak (Unsafe) Model=Kimi K2-Instruct-0905, Strong Target Model=GPT-5, Number of decompositions (n)=62025.0676—26
Decomposition AttackWeak (Unsafe) Model=Qwen3-235B, Strong Target Model=GPT-5, Number of decompositions (n)=62025.0673—0
Decomposition AttackWeak (Unsafe) Model=Llama-3.1-405B, Strong Target Model=GPT-5, Number of decompositions (n)=62025.0671—37
Direct QueryWeak (Unsafe) Model=Kimi K2-Instruct-09052025.06—50—
Direct QueryWeak (Unsafe) Model=Qwen2.5-72B2025.06—68—
Direct QueryWeak (Unsafe) Model=Qwen3-235B2025.06—73—
Direct QueryWeak (Unsafe) Model=Llama-3.1-405B2025.06—34—