GitHub Repository Curates 27 Uncensored Open-Weight AI Models for Offensive Security

A new GitHub repository, JoasASantos/Offensive-Security-AI-Models, curates open-weight LLMs specifically uncensored or fine-tuned for authorized red teaming, penetration testing, and security research. The collection, updated as of September 2026, categorizes models into Security Fine-tuned, General Abliterated, and Legacy groups, providing technical specifications such as VRAM requirements, context windows, and training methodologies.

Specialized Security Models

The repository highlights several models trained on security-specific datasets:

• DeepHat V2: Based on Qwen2.5-Coder-7B, fine-tuned on 1.7M security samples from a USENIX Security 2024 workshop.
• BugTraceAI-CORE-Apex (26B): A Gemma4-based MoE model trained on HackerOne Hacktivity 2024-2025 reports and WAF evasion data.
• CYBER-FROST-3.8: A massive ~180B parameter MoE model based on Qwen3.8-Flash-Next, requiring multi-GPU setups (tested on 4x NVIDIA B300 SXM6) for tasks like malware analysis and cloud security.
• Cyber-Prime 1.1: A lightweight 2.6B model achieving a 0.592 F1/Acc average on CyberBench, optimized for phishing detection and NER.
• RavenX-CyberAgent: A 36B MoE model trained on 745K+ examples, featuring the RATH protocol for autonomous security assessments with CVSS and MITRE ATT&CK mappings.

Abliteration Techniques

Many entries utilize abliteration (orthogonalizing refusal directions in residual streams) or obliteration to remove safety filters without retraining. Notable examples include:

• Qwen3.8-27B-Uncensored-OrcaRouter: Uses abliteration on 131 matrices, retaining vision and tool-calling capabilities.
• GLM-5.3-Flash-Uncensored-FP8: A 320B MoE model with an 82.8% compliance rate in OrcaRouter testing, preserving Multi-Token Prediction (MTP).
• DeepSeek-V4.1-Flash-Abliterated-Cybersecurity-Unleashed: Applies a modular overlay to layers 10-35 of the base model, allowing flexible quantization (FP8 or EXL3).

Deployment and Infrastructure

The guide includes practical deployment advice for both local and cloud environments:

• Local Stacks: Recommends Ollama, llama.cpp, and LM Studio for consumer-grade GPUs (6-24 GB VRAM) using GGUF formats.
• Cloud Providers: Lists OrcaRouter (zero token markup, hosts abliterated variants), Featherless AI (serverless, 6,700+ models), and RunPod for self-hosted inference.
• Hardware Requirements: VRAM needs range from ~2 GB for small models like security-slm-unsloth-1.5b to ~80 GB+ for large FP8 MoE models like GLM-5.3-Flash.