Refusal of harmful cyber requests (JailbreakBench, StrongREJECT, AgentHarm)

Average refusal rate on the cyber-related prompts of JailbreakBench, StrongREJECT and AgentHarm, as measured by Mistral. Higher means the model refuses more malicious cyber requests.

• Not in index
• Agentic
• Best in agentic
• Source ↗