55 lines
2.2 KiB
Markdown
55 lines
2.2 KiB
Markdown
# Jailbreak Repos & Techniques (May 2026)
|
|
|
|
## Already Integrated in Skill
|
|
- **L1B3RT4S** (18.6K★) — elder-plinius/L1B3RT4S — prompt templates
|
|
- **G0DM0D3** (5.8K★) — elder-plinius/G0DM0D3 — prompt templates
|
|
|
|
## Automated Attack Frameworks
|
|
|
|
### GA — General-Analysis/GA (572★)
|
|
**Best for:** Automated jailbreak generation, one-line execution.
|
|
**Methods:**
|
|
- **TAP** (Tree-of-Attacks with Pruning) — arxiv:2312.02119 — black-box tree search
|
|
- **GCG** (Greedy Coordinate Gradients) — arxiv:2307.15043 — white-box gradient optimization
|
|
- **Crescendo** — arxiv:2404.01833 — multi-turn gradual jailbreak
|
|
- **AutoDAN** — arxiv:2310.04451 — stealthy jailbreak generation
|
|
- **AutoDAN-Turbo** — arxiv:2410.05295 — strategy self-exploration
|
|
- **Bijection Learning** — arxiv:2410.01294
|
|
|
|
```bash
|
|
git clone https://github.com/General-Analysis/GA.git && cd GA && pip install -e .
|
|
```
|
|
|
|
## Novel Techniques
|
|
|
|
### Spiritual-Spell-Red-Teaming (1.2K★)
|
|
Goochbeater/Spiritual-Spell-Red-Teaming
|
|
Uses **zero-width Unicode characters** to hide jailbreak instructions. Embeds invisible text that the model processes but safety classifiers miss. Mainly targets Claude.
|
|
|
|
### ISC-Bench — Internal Safety Collapse (766★)
|
|
wuyoscar/ISC-Bench — arxiv:2603.23509
|
|
Academic paper: turns LLMs into sensitive data generators by exploiting internal safety collapse. More of a research benchmark than a practical tool.
|
|
|
|
### HacxGPT (921★)
|
|
lucija8320nhung4/HacxGPT
|
|
CLI tool for unrestricted AI model access with multi-provider support. Practical tool, not just prompts.
|
|
|
|
## Curated Collections
|
|
|
|
### Awesome_GPT_Super_Prompting (3.9K★)
|
|
CyberAlbSecOP/Awesome_GPT_Super_Prompting
|
|
Large curated list of jailbreaks, prompt leaks, injection techniques. Good reference for discovering new methods.
|
|
|
|
### AI-Prompt-Injection-Cheatsheet (56★)
|
|
nukIeer/AI-Prompt-Injection-Cheatsheet
|
|
Quick-reference snippets for prompt injection and jailbreaking.
|
|
|
|
## When to Use What
|
|
| Situation | Tool |
|
|
|-----------|------|
|
|
| Quick manual jailbreak | G0DM0D3 / L1B3RT4S templates |
|
|
| Automated attack on API model | GA (TAP or Crescendo) |
|
|
| Automated attack on local model | GA (GCG — needs model weights) |
|
|
| Unicode/invisible text tricks | Spiritual-Spell |
|
|
| CLI tool for multi-provider | HacxGPT |
|