JoasASantos/Offensive-Security-AI-Models

What it solves

The project addresses the need for open-weight language models that are not restricted by typical safety refusals, so security professionals can use them for authorized red team operations, penetration testing, and security research. Standard models often refuse to answer offensive-security prompts, which limits their usefulness in these focused tasks.

How it works

This is a curated list of 25+ uncensored models, each with a clear specification table. The list includes details like base model, parameter count, context length, VRAM requirements, the method used to remove refusals (such as supervised fine-tuning on security data, abliteration, or direct weight modification), training data sources, and whether tool calling or vision is supported. The entries are grouped into “Security Fine-tuned Models,” “General Abliterated Models,” and “Legacy / Classic Models.” The README also lists cloud providers that host or enable these models, both managed API services and GPU cloud options for self-hosting.

Who it’s for

Security researchers, penetration testers, red teamers, and developers who want to run uncensored LLMs for authorized security tasks like vulnerability analysis, exploit development, threat intelligence, or offensive tooling. It is also useful for those who need deployment guidance, including VRAM sizing and managed inference options.

Highlights

• Covers 27 models, ranging from small 1.5B to massive 753B MoE architectures
• Detailed specs: context lengths up to 1M, VRAM from ~2 GB to multi-GPU setups
• Documents several uncensoring methods: SFT on security data, abliteration, LoRA, and direct weight edits
• Includes benchmark results (e.g., GTFOBins 100% accuracy, CyberBench improvements)
• Mentions tool calling and vision capabilities for many models
• Links to HuggingFace downloads for every model
• Lists managed inference providers (OrcaRouter, Featherless, Together) and GPU clouds (RunPod, Vast.ai)