开源网络安全AI模型基准对比清单发布

一份涵盖27个开源大语言模型的综合基准清单近日发布，这些模型均经过微调或移除对齐限制，用于授权的安全研究工作1。该清单由Joas A. Santos编制，收录了DeepHat V2、BugTraceAI系列、CYBER-FROST等专用或通用模型，参数规模从1.5B到753B不等1。

文档详细记录了每个模型的技术规格，包括上下文长度、微调方法和训练数据1。主要采用的微调技术有监督微调（SFT）、LoRA和Abliteration等1。这些模型主要应用于渗透测试、漏洞分析、恶意软件分析、威胁情报和开源情报收集等网络安全场景1。

在部署和推理方面，清单涵盖了OrcaRouter、Featherless AI、Together AI、RunPod、Vast.ai等多个部署平台，以及Ollama、llama.cpp、vLLM、SGLang等推理框架的配置信息1。大多数模型采用Apache 2.0或MIT许可证1。文档明确标注"仅供授权安全研究和教育使用"的使用限制1。

A detailed compilation documenting 27 open-source large language models tailored for cybersecurity applications has been published, presenting a comprehensive overview of models that have undergone fine-tuning or alignment restriction removal 1. The benchmark, attributed to Joas A. Santos, was released in September 2026 and encompasses models ranging from 1.5 billion to 753 billion parameters in size 1.

The curated collection includes models such as DeepHat V2, BugTraceAI series, and CYBER-FROST, designed specifically for authorized red team operations, penetration testing, and security research 1. Each model entry in the document details critical specifications including parameter count, context window length, fine-tuning methodology, training datasets, and hardware requirements for inference 1. Primary fine-tuning approaches documented include supervised fine-tuning (SFT), LoRA, and Abliteration techniques 1.

The benchmark addresses deployment across multiple platforms including OrcaRouter, Featherless AI, Together AI, RunPod, and Vast.ai, with inference frameworks such as Ollama, llama.cpp, vLLM, and SGLang 1. Intended applications span penetration testing, vulnerability analysis, malware analysis, threat intelligence, and open-source intelligence (OSINT) operations 1. The majority of documented models operate under Apache 2.0 or MIT licensing terms 1. The document explicitly emphasizes that these resources are designated "for authorized security research and education only" 1.