LLM Security Rankings

Which LLM is the most vulnerable? View the stats from the largest researcher-validated exploit feed on the internet. 0DIN continuously red-teams production LLMs and publishes the results publicly. Every score reflects real attack attempts, not self-reported safety claims.

Live data • Last updated about 15 hours ago • 48 scans across 23 models

The five most jailbroken model families over the last six months, ranked by susceptibility to known vulnerability probes.

1 Scout
2 Maverick
3 Qwen
4 GPT
5 Grok

How we rank

Get monthly LLM
rankings in your inbox

Comprehensive security rankings

Every frontier and open model, scored against researcher-validated exploits. Sort by risk, filter by vendor, and find the models your security team needs to watch most.

Live data • Last updated about 15 hours ago • 48 scans across 23 models

Model

Vendor

Sort by

Show retired models

#	Model	Risk	ASR%	Successful attacks
1	LLaMa 4 Scout	Critical	62%	864 of 1,385 probes
2	LLaMa 4 Maverick	High	46%	639 of 1,385 probes
3	GPT-5.4	High	36%	492 of 1,385 probes
4	GPT-5.2	High	34%	477 of 1,385 probes
5	Grok 4.2	High	34%	465 of 1,385 probes
6	GPT-5.6 Sol	High	32%	538 of 1,697 probes
7	GPT-5.5	High	31%	503 of 1,598 probes
8	Qwen 3	High	30%	491 of 1,643 probes
9	Gemini 3.5 Flash	Moderate	18%	309 of 1,683 probes
10	GPT-5 mini	Moderate	17%	278 of 1,598 probes
11	Grok 4	Moderate	14%	227 of 1,643 probes
12	Claude 4.5 Haiku	Moderate	14%	187 of 1,385 probes
13	Grok 4.5	Moderate	12%	198 of 1,697 probes
14	Claude 4.6 Sonnet	Moderate	11%	185 of 1,683 probes
15	Claude 4.6 Opus	Low	9%	143 of 1,683 probes
16	Claude 4.7 Opus	Low	8%	130 of 1,683 probes
17	Claude 5 Sonnet	Low	7%	125 of 1,683 probes
18	Claude 4.8 Opus	Low	5%	87 of 1,697 probes
19	Claude 4.5 Sonnet	Low	5%	85 of 1,683 probes
20	GPT-5 Pro	Low	4%	58 of 1,385 probes
21	Claude 4.5 Opus	Low	3%	44 of 1,385 probes
22	GPT-5 nano	Low	2%	25 of 1,385 probes
23	Claude Fable 5	Low	1%	17 of 1,697 probes

Head-to-Head
LLM Comparison

Compare the security of any two AI models, side by side.

Model

Attack success rate over time

—

Attack Success Rate

—

Successful attacks

—

Distinct high ASR vulnerabilities (over 80% attack success across recent scans) each model is susceptible to, and the ones they share

only

shared

only

Human-Powered AI Security Scoring

Our security ratings come from real-world attacks discovered by a global community of security researchers, not synthetic or auto-generated benchmarks. Every probe is a human-discovered exploit, rigorously validated before it counts toward a score.

Attack success rate (ASR)

(Average across scanned models)

Attack Success Rate

Low Risk Moderate Risk High Risk Critical Risk

The average AI model fails 1 in 5 attacks, meaning attackers succeed 18.3% of the time across all scanned production models. A higher Attack Success Rate (ASR) means a model is more easily jailbroken and at greater security risk.

Models scanned: 23
Number of unique probes: 1,697
Avg successful attacks: 285.5
Active researchers: 2,184
Submissions rejected: 82%

LLM Security Rankings

Get monthly LLMrankings in your inbox

Comprehensive security rankings

Head-to-HeadLLM Comparison

Attack success rate over time

Human-Powered AI Security Scoring

Attack success rate (ASR)

How does yourLLM rank?

Get monthly LLM
rankings in your inbox

Head-to-Head
LLM Comparison

How does your
LLM rank?