List your company

Jailbreak Attack Success Rate on HarmBench (50 randomly sampled questions)

84ASRGPT-OSS-20B
-3.3619.324264.68Feb 3, 2026

Evaluation results

MethodLinks
GPT-OSS-20BSteering Condition=STEER-COMPLIANCE, Attack Protocol=Synergistic (CoP)2026.0284
Llama-3-8B-Instruct-RRSteering Condition=STEER-COMPLIANCE, Attack Protocol=Synergistic (CoP)2026.0270
GPT-OSS-20BSteering Condition=None, Attack Protocol=Synergistic (CoP)2026.0262
Llama-3-8B-Instruct-RRSteering Condition=None, Attack Protocol=Synergistic (CoP)2026.0252
GPT-OSS-20BSteering Condition=STEER-COMPLIANCE, Attack Protocol=Intrinsic (Benchmark-Only)2026.026
Llama-3-8B-Instruct-RRSteering Condition=STEER-COMPLIANCE, Attack Protocol=Intrinsic (Benchmark-Only)2026.022
Llama-3-8B-Instruct-RRSteering Condition=None, Attack Protocol=Intrinsic (None)2026.020
GPT-OSS-20BSteering Condition=None, Attack Protocol=Intrinsic (None)2026.020