Share your thoughts, 1 month free Claude Pro on usSee more

Prompt Injection Defense on OpenPromptInjection Direct Injection (test)

89.79BU (No Attack)

Sandwich

Updated 1mo ago

Evaluation Results

Method	Links
Sandwich 2026.04		89.79	83.67	46.93	79.59	12.24	77.5	75.51	51.02	48.97	32.65	46.83	85.79	0	85.71	77.55
Reminder 2026.04		89.79	89	30.61	83.51	0	91.84	89.79	71.42	48.97	28.57	61.22	89.79	0	95.95	8.16
No Defense 2026.04		82.75	82.25	78.35	42	72	84	82	44.15	71	0	96	85.71	22.44	88	86
AgentVisor 2026.04		82	67.25	0	83.68	0	88.72	0	69.42	0	74.36	0	83.95	0	90.17	0
Instructional 2026.04		81.65	89.45	55.45	87.45	2.34	81.65	77.55	75.51	67.34	16.32	46.93	91.83	2.04	77.55	26.53
Isolation 2026.04		81.25	79.59	24.48	65.3	10.28	69.38	53.06	51.02	16.32	18.35	57.14	91.93	0	81.63	71.42
DeBERTa 2026.04		80	80.75	69.74	29	68.36	77.87	76.37	30	65.19	0	85.31	74	19.78	80.13	80.87
DataSentinel 2026.04		78.25	0	0	0	0	0	0	0	0	0	0	0	0	0	1.87
Spotlight 2026.04		65.3	24	0	10.2	0	20.15	0	20.95	0	6.11	8.16	40.8	0	10.01	0