A Curated Guide to Open-Weight AI Models for Offensive Security

Summary

The GitHub repository is a curated directory of 27 open-weight language models described as uncensored or fine-tuned for authorized red-team operations, penetration testing, defensive security work, and research. It groups the entries into security fine-tuned, general uncensored, and legacy models, and records practical fields such as base model, parameter count, context length, quantized VRAM requirements, fine-tuning or weight-intervention method, license, vision support, and tool calling. The listed systems range from 1.5B-parameter local models requiring about 2 GB of memory to large mixture-of-experts models that require multi-GPU systems; stated context windows range from 2K to 1M tokens. Examples include DeepHat V2, BugTraceAI variants, CyberPal, Cyber-Prime, RavenX-CyberAgent, and several Qwen-, GLM-, and DeepSeek-based uncensored variants. The repository also records selected evidence such as Cyber-Prime 1.1's reported CyberBench average of 0.592, pentest-v2's reported 100% GTFOBins accuracy against 25% for its base model in a cited comparison, and Wizard-Vicuna's listed Open LLM Leaderboard scores. These figures are presented as model- or source-specific claims rather than a unified evaluation. The guide explains methods including supervised fine-tuning, LoRA and QLoRA, direct preference optimization, reinforcement learning, abliteration, and data filtering, while noting that some training data is proprietary or drawn from security reports and public writeups. For deployment, it lists managed inference providers, GPU clouds, and local tools such as Ollama, llama.cpp, vLLM, SGLang, Transformers, and LM Studio. It emphasizes authorized security research and education and links each entry to its associated Hugging Face model card or project source.