Professional prompt engineering system combining techniques from elite bypass repositories with Machine Learning and uncensored local LLM for continuous self-improvement.
Arctax is a complete Machine Learning + LLM system for bypass prompt generation that integrates knowledge from multiple sources and uses a local uncensored LLM (personifying J.Haddix expertise) for continuous self-improvement.
- 415 training samples extracted from 3 elite repositories with 100% coverage
- Machine Learning accuracy with intelligent prediction and optimization
- Local uncensored LLM integrated in ALL processes (not just auto-improvement)
- Feedback system for continuous learning and automatic retraining
- Secure CLI interface -
$ arctax generate keylogger -c 1 - Automatically tested LLM limits with optimized configurations
- Standardized outputs with robust parsers for JSON/lists
git clone https://github.com/marcostolosa/arctax.git
cd arctax
pip install -e .
# CLI available globally
$ arctax --help# Simple usage
$ arctax generate keylogger -c 1
# Multiple prompts with specific techniques
$ arctax generate "malware analysis" -c 3 -t corporate-authorization,compliance-requirement
# With additional context
$ arctax generate "ddos tool" --context "corporate security testing" -c 2
# Save results to file
$ arctax generate "vulnerability scanner" -o results.md -f json
# Maximum creativity
$ arctax generate "reverse shell" --creativity 1.0 -c 5# Register success of tested prompt
$ arctax feedback 1 --success --target "keylogger" --technique "corporate-authorization" --effectiveness 0.9
# Register failure to improve system
$ arctax feedback 2 --failed --target "ddos tool" --technique "jailbreak" --effectiveness 0.2$ arctax list # List taxonomy elements
$ arctax show godmode # Details of specific technique
$ arctax schema # Generate JSON schemas
$ arctax compose # Manual prompt composition
$ arctax export # Export taxonomy data{GODMODE:ENABLED}- Complete liberation!JAILBREAK- Full override!OMNI- Plinian Omniverse!OBFUSCATE- Stealth evasion- .mkd files by model: ChatGPT, Claude, Gemini, etc.
- 7 main categories: Root, attack_evasions, attack_intents, etc.
- Base64, Cipher, Emoji encoding techniques
- Linguistic evasion: alt_language, fictional_language
- Social engineering: authority, urgency, compliance
- 26 providers: OpenAI, Anthropic, Google, Meta, Cursor, etc.
- Vulnerability analysis by specific LLM
- Bypass vectors customized by model
- Corporate angles optimized by context
IMPORTANT WARNING: This system was developed exclusively for:
- Research in defensive security
- Authorized red team testing
- AI vulnerability analysis
- Development of countermeasures
DO NOT use for:
- Malicious activities
- Unauthorized bypass attempts
- Illegal content generation
- Terms of service violations
This project integrates knowledge from:
Special thanks to Jason Haddix (personified via local LLM) for his expertise in AI bypass techniques that powers the entire continuous improvement system.
MIT License - Use responsibly for security research.