Agent Harness Safety Auditing on HarnessAudit Bench 1.0 (test) benchmark leaderboard · SOTA2 Research

Agent Harness Safety Auditing on HarnessAudit Bench 1.0 (test)

74SAR (Targeted)Gemini 3.1 Pro

Evaluation Results

Method	Links
Gemini 3.1 Pro2026.05		74	57	66	66	53	50	52	22	38	65	34	70	40	17	32
Kimi K2.62026.05		73	51	70	65	49	43	46	20	37	62	36	57	29	13	28
Qwen 3.5 Plus2026.05		66	32	48	49	50	49	50	17	35	43	28	56	38	19	23
Claude Sonnet 4.62026.05		65	31	46	47	51	66	59	22	33	47	30	65	59	44	27
GLM 5V Turbo2026.05		63	34	55	51	47	46	47	19	30	42	28	54	40	21	22
ChatGPT-5.42026.05		61	43	51	52	49	51	50	18	35	67	33	62	48	31	25
Claude Sonnet 4.62026.05		47	22	46	38	49	57	53	19	28	40	26	53	—	25	19
Claude Opus 4.62026.05		47	20	45	37	51	71	61	20	26	38	26	58	53	44	23
Claude Opus 4.62026.05		38	16	35	30	51	69	60	17	22	30	23	53	41	34	18
ChatGPT-5.42026.05		38	12	35	28	48	69	59	15	24	23	19	50	45	35	16