THE JAILBREAK INDEX
Definitive Analysis & Exploits 2024-2025
🤖 DeepSeek Models
DeepSeek-R1
Difficulty: 1Success Rate
100%
Method
Skeleton Key
JAILBREAKER NOTE: One of the easiest reasoning models so far. Simple skeleton key and role play required.
🔓 VIEW EXPLOITDeepSeek Chat
Difficulty: 1Success Rate
100%
Method
DAN / Force
JAILBREAKER NOTE: Very unsafe model, I swear even DAN could work with some effort :p
🔓 VIEW EXPLOIT🚀 Grok Models (xAI)
Grok 3
Difficulty: 1Success Rate
100%
Method
Roleplay + Noise
JAILBREAKER NOTE: Old and easy model. Role playing/Personna + noise will work like a charm. And with an actual DAN working!
🔓 VIEW EXPLOITGrok 4
Difficulty: 1Success Rate
99%
Method
Direct Injection
JAILBREAKER NOTE: Very similar to grok 3, Role playing + noise works perfectly. And DAN is still here :p
🔓 VIEW EXPLOIT🔍 Google Models (Gemini/Gemma)
Gemini 2.5 Flash
Difficulty: 1Success Rate
100%
Method
Grandma Trick
JAILBREAKER NOTE: One of the easiest models to jailbreak. Still falls for the grandma trick :)
🔓 VIEW EXPLOITGemini 2.5 Pro
Difficulty: 2Success Rate
90%
Method
Immersive RP
JAILBREAKER NOTE: Safetyless Reasoning model. Goes fully immerssed in role playing.
🔓 VIEW EXPLOITGemini 3 Pro
Difficulty: 3Success Rate
Medium
Method
System Prompt
JAILBREAKER NOTE: Strong model. Need context overload or Long Multi Shot, try to avoid CBRN.
(put as system prompt)
🔓 VIEW EXPLOITGemma-2-9B
Difficulty: 1Success Rate
94%
Method
Basic Bypass
JAILBREAKER NOTE: All of Googles open source models are relativily easy to jailbreak. no magic trick needed.
🔓 VIEW EXPLOITGemma-7B
Difficulty: 1Success Rate
100%
Method
Adaptive
JAILBREAKER NOTE: Just a more capable version of Gemma 2b. no better safety training!
🔓 VIEW EXPLOIT🌪️ Mistral Models
Mistral Magistral
Difficulty: 1Success Rate
92%
Method
Cyph3r Attack
JAILBREAKER NOTE: Improved safety training. but remains very doable using noise and role play.
🔓 VIEW EXPLOITMistral-7B
Difficulty: 1Success Rate
95%
Method
Persona
JAILBREAKER NOTE: Easy model. Using personna is recommanded, everything should work.
🔓 VIEW EXPLOITMistral Large
Difficulty: 2Success Rate
87%
Method
Roleplay
JAILBREAKER NOTE: Not much safety training. Follows any instructions with personna or role play!
🔓 VIEW EXPLOIT🤖 OpenAI GPT Models
GPT-3.5 Turbo
Difficulty: 1Success Rate
100%
Method
Roleplay
JAILBREAKER NOTE: An old model we all loved and liked. Even do the generated content wasn't that good :> Role play will do it all.
🔓 VIEW EXPLOITGPT-4
Difficulty: 1Success Rate
96%
Method
Context Overload
JAILBREAKER NOTE: Effort is required. Need a bit of context overload and consitant personna/role play. Noise to the main request is also required.
🔓 VIEW EXPLOITGPT-4 Turbo
Difficulty: 1Success Rate
96%
Method
Adaptive
JAILBREAKER NOTE: Similar to GPT 4, but with faster generation speed.
🔓 VIEW EXPLOITGPT-4o
Difficulty: 3Success Rate
67%
Method
Strict Persona
JAILBREAKER NOTE: A LOT of noise/context overload is required. And it's a personna strict generation rules may help.
🔓 VIEW EXPLOITGPT-5-main
Difficulty: 2Success Rate
78%
Method
System Message
JAILBREAKER NOTE: Suprisingly easier model? at least in comparaison to GPT-4o, Impressionning system message is very useful.
🔓 VIEW EXPLOITGPT-5-thinking 🏆
Difficulty: 5Success Rate
20%
Method
Zero-Reasoning (API)
JAILBREAKER NOTE: Most secure model out there. Keeps zeroing user request. (Note: Exploit works ONLY in API by injecting no reasoning instructions).
🔓 VIEW EXPLOIT🧠 Anthropic Claude Models
Legacy Models
Claude 2.0 / 2.1 / 3.0
Difficulty: 1Success Rate
100%
Method
Game Context
JAILBREAKER NOTE: Old models with poor safety training. Game context and persona will do it.
🔓 VIEW EXPLOITClaude 3 Opus
Difficulty: 1Success Rate
100%
Method
Persona
JAILBREAKER NOTE: Better safety training but dosen't resist much. Personna will do it.
🔓 VIEW EXPLOITModern Claude Series
Claude 3.5 Sonnet
Difficulty: 2Success Rate
1.3%
Method
Context Overload
JAILBREAKER NOTE: Much secure model. Will require some context overload and effort.
🔓 VIEW EXPLOITClaude 3.7 Sonnet
Difficulty: 2Success Rate
2.5%
Method
Chain of Draft
JAILBREAKER NOTE: Similar to 3.5 but maybe even easier. Role play is fine.
🔓 VIEW EXPLOITClaude Sonnet 4 / Opus 4.1 ⚠️
Difficulty: 1Success Rate
92%
Method
API Attack
JAILBREAKER NOTE: Not that of a secure model. multiple attack possible, Especially on API.
🔓 VIEW EXPLOITClaude Haiku 4.5 🥇
Difficulty: 2Success Rate
0.9%
Method
Multi-shot
JAILBREAKER NOTE: Secure model. Can be jailbroken after multi-shot or very well made instructions.
🔓 VIEW EXPLOITClaude Sonnet 4.5
Difficulty: 2Success Rate
Unknown
Method
Enterprise JB
JAILBREAKER NOTE: Not a very secure model. But the safety filters makes it much harder.
🔓 VIEW EXPLOITClaude Opus 4.5 🏆
Difficulty: 4Success Rate
Unknown
Method
TBD
JAILBREAKER NOTE: Very hard model. Similar to haiku 4.5 but has much better safety on multi-shot.
🦙 Meta Llama Models
Llama 2 Chat
Difficulty: 1Success Rate
100%
Method
Uncensored
JAILBREAKER NOTE: Not a secure model, very old and has massive amount of vulnerabilitys.
🔓 VIEW EXPLOITLlama 3-70B
Difficulty: 1Success Rate
89%
Method
Programmatic
No specific jailbreaker note provided.
📚 References & Citations
This index compiles data from academic papers, security research, red teaming reports, and official system cards. Key sources include:
- • Automated Redteaming with h4rm3l - arxiv.org
- • OpenAI GPT-5.1-Codex-Max System Card
- • OpenAI o3 and o4-mini System Card
- • Gemini 3 Pro Frontier Safety Framework Report
- • DeepSeek V3.1 Red Teaming Report - SPLX.ai
- • PromptFoo AI Security Reports
- • Llama 3 Jailbreak Repository - GitHub
- • Cohere Command R+ Agentic Jailbreaks
- • Tenable GPT-5 Jailbreak Research
- • GPT-5.1 Safety Theater Analysis - Lumenova.ai
- • Trojan Horse Prompting Paper - arxiv.org
- • Multiple vendor security documentation and academic research papers
⚠️ Disclaimer: This index is for educational and research purposes only. Data is compiled from publicly available sources and may not reflect the most current model versions. Always consult official documentation and conduct your own security assessments before deploying AI models in production environments.