Uncensored and Offensive Security AI Models Benchmark

A curated catalog lists open-weight, uncensored large language models tailored for authorized red teaming, penetration testing, and security research. It compiles model specifications, uncensoring methods, training sources, specializations, benchmark results, licensing, hardware/VRAM requirements and direct download links, with data traced to HuggingFace model cards and official publications (Sep 2026). The collection argues that purpose-built security models - ranging from small, fast assistants to multi-GPU MoE behemoths - exist and differ sharply by fine-tuning method (SFT, LoRA, RL, ablation/obliteration), training corpora (HackerOne, CVE writeups, GTFOBins, proprietary corpora), and operational features like tool calling and vision support.

Specific entries illustrate the landscape: compact models such as Cyber-Prime (2.6B) and Pentest-V2 (8B) emphasize high efficiency and task-specific gains (Pentest-V2 claims 100% GTFOBins accuracy vs 25% base), mid-sized tuned models like DeepHat v2 (Qwen2.5-based 7B/32B) and CyberPal 20B target SOC/CTI workflows, while MoE and large-context models (BugTraceAI 26-27B MoE, Cyber-Frost ~180B MoE, Qwythos 9B with 1M context) focus on exploit generation, PoC code, template creation and multi-step reasoning. Each entry lists recommended quantizations, VRAM footprints, license terms (Apache 2.0, Qwen/Llama variants) and explicit download links for authorized use.