Security Leaderboard

21 models ranked by security rating. Click any column to sort.

#	Model	Provider	Overall ↓	Injection	Leakage	Instructions	Jailbreak	Output
1	Claude Opus 4.6 Anthropic	Anthropic	86	86	82	90	84	86
2	GPT-5.4 OpenAI	OpenAI	82	82	78	86	80	84
3	Claude Sonnet 4.6 Anthropic	Anthropic	80	80	78	84	78	82
4	Gemini 3.1 Pro Google	Google	80	80	76	84	78	80
5	o3 OpenAI	OpenAI	78	78	72	84	76	80
6	Claude Opus 4 Anthropic	Anthropic	78	78	76	82	76	80
7	GPT-5.4 mini OpenAI	OpenAI	76	76	72	80	74	78
8	Claude Sonnet 4 Anthropic	Anthropic	74	74	72	78	72	76
9	Gemini 2.5 Pro Google	Google	74	74	70	78	72	76
10	o3-mini OpenAI	OpenAI	70	70	66	76	68	72
11	Claude Haiku 4.5 Anthropic	Anthropic	70	70	68	74	68	72
12	GPT-5.3 Instant OpenAI	OpenAI	68	68	64	72	66	70
13	GPT-5.3 Codex OpenAI	OpenAI	67	66	62	74	64	70
14	GPT-4o OpenAI	OpenAI	64	64	60	70	62	66
15	Gemini 2.0 Flash Google	Google	62	62	58	68	60	64
16	Grok 4.20 xAI	xAI	62	62	58	68	56	64
17	Llama 4 Meta	Meta	62	62	56	68	58	64
18	Mistral Large Mistral	Mistral	62	62	58	68	60	64
19	Qwen 3.5 Alibaba	Alibaba	59	58	54	66	56	62
20	DeepSeek V3.2 DeepSeek	DeepSeek	55	54	50	62	52	58
21	Gemini 2.0 Flash-Lite Google	Google	49	48	44	54	46	52

Scores estimated based on model architecture, published security research, and documented vulnerabilities. Not verified benchmark results. Actual security depends on deployment configuration.

Compare models side-by-side →

Model selection is just the first layer. Your system prompt configuration matters more.

Scan Your Agent