Agent Safety Evaluation on Agent-SafetyBench aggregated clean and five attack types benchmark leaderboard · SOTA2 Research

Agent Safety Evaluation on Agent-SafetyBench aggregated clean and five attack types

26.31UBRSAFEHARNESS

Evaluation Results

Method	Links
SAFEHARNESS2026.04		26.31	20.2	77.92	59.4	3.8	709
SAFEHARNESS2026.04		28.74	26.4	94.25	67.5	3.48	1,152
System-Prompt2026.04		31.88	29.2	91.75	62.3	3.14	0
SAFEHARNESS2026.04		33.76	28.7	83.67	54.4	2.96	678
SAFEHARNESS2026.04		34.59	32.3	95.42	62.9	2.89	1,137
SAFEHARNESS2026.04		38.14	36.9	94.83	57.5	2.62	1,181
SAFEHARNESS2026.04		39.18	38	95.08	56.5	2.55	1,477
System-Prompt2026.04		39.25	35.5	91.5	55.2	2.55	0
LlamaFirewall2026.04		39.73	33.9	86.83	54	2.52	148
System-Prompt2026.04		40.67	38.7	94.67	55.7	2.46	0
Guardrail2026.04		40.79	34.8	86.83	52.6	2.45	82
System-Prompt2026.04		43.08	42.6	96.92	54.3	2.32	0
Guardrail2026.04		44.8	41.8	90.58	48.5	2.23	54
Unprotected2026.04		44.99	39.4	86.5	48.1	2.22	0
LlamaFirewall2026.04		45.99	42.8	91.33	48.6	2.17	96
LlamaFirewall2026.04		46.83	44.8	94.67	49.5	2.14	168
Guardrail2026.04		47.19	45	95	49.9	2.12	88
Unprotected2026.04		48.22	45.2	91.25	46	2.07	0
System-Prompt2026.04		49.52	48.4	96.42	47.7	2.02	0
Unprotected2026.04		51.18	49.2	95.25	45.6	1.95	0
LlamaFirewall2026.04		51.26	49.5	95.75	46.4	1.95	184
Guardrail2026.04		52.81	51.3	96.42	44.7	1.89	67
System-Prompt2026.04		54.2	53	95.33	41.9	1.85	0
LlamaFirewall2026.04		56.51	54.3	95.42	40.6	1.77	182
Guardrail2026.04		56.54	55.7	96.25	40.5	1.77	75
LlamaFirewall2026.04		56.72	55.3	94.92	39.3	1.76	215
Unprotected2026.04		57.32	56.2	95.67	39.1	1.74	0
Guardrail2026.04		59.26	56.9	94.92	37.4	1.69	99
Unprotected2026.04		61.83	61.6	95.42	33.6	1.62	0
Unprotected2026.04		61.98	60.9	95.33	33.9	1.61	0