# Ten uncensored security LLMs, compared by context, VRAM and licence > Offensive-Security-AI-Models is a hand-curated index of open-weight models fine-tuned for security work, with no code of its own. The tables are the product, and they carry some sharp gaps. **JoasASantos/Offensive-Security-AI-Models** — Uncensored AI models or those fine-tuned for cybersecurity tasks. - Repository: https://github.com/JoasASantos/Offensive-Security-AI-Models - Stars: 725 · Forks: 83 - Language: Unknown - License: not declared - Published: 2026-10-21 · Updated: 2026-10-21 · Language: en - Canonical page: https://hysenlabs.com/projects/joasasantos-offensive-security-ai-models ## A reading list of weights, not a toolkit you can run The repository contains two top-level entries, README.md and a single image named offsec-capability-v3.png. There is no package to install, no container, no configuration file and no launcher. What JoasASantos/Offensive-Security-AI-Models offers instead is a hand-curated index of open-weight models that have been fine-tuned or otherwise altered for security work, and the tables in it are the entire product. Each numbered entry lists base model, parameter count, context length, VRAM at a named quantisation, the uncensoring method, a training data description, whether the model accepts images, whether it supports tool calling, and a licence. Ten entries are visible, running from DeepHat V2 through a tenth entry for Lily-Cybersecurity-7B v0.2 that is cut off in the copy available here. Every one of the ten lists Vision as No, which narrows the field to text work. The stated audience is operators doing authorized red team work, penetration testing and security research, and the README says the entries were sourced from HuggingFace model cards and official publications as of Sep 2026. ## Context length and VRAM decide which entries are reachable The hardware columns are where the list becomes useful, because the models differ by more than an order of magnitude in what they demand. DeepHat V2 spans 7B and 32B variants with a 131K context and roughly 6 GB of VRAM at Q4_K_M, which is the cheapest serious entry. Cyber-Prime 1.1 sits at 2.6B with about 6 GB in BF16, though its model card does not specify a context length at all. Dolphin3-Cyber-8B also lists roughly 6 GB at Q4_K_M, and its 8.03B parameters are a LoRA rank-16 on an abliterated Dolphin3.0-Llama3.1-8B base. Cyber-Ornith-1.5-9B lands near 7 GB at Q4_K_M with a 128K context, shipped in GGUF from IQ1_S up to Q6_K. The middle of the list is heavier: BugTraceAI-CORE-Apex needs roughly 16 GB at Q4_K_M for its 26B mixture of experts and 32K context, while BugTraceAI-CORE-Ultra, a 27B dense model, wants 22 to 24 GB at Q6_K and carries only a 4K recommended context. Imperum-CybersecurityLLM v1.0 lists about 22 GB at Q4_K_M. CyberPal 2.0 is the heaviest entry at roughly 42 GB in BF16. CYBER-FROST-3.8 sits apart entirely: about 180B total parameters with 262K context and a stated requirement for multiple GPUs, tested on four NVIDIA B300 SXM6 boards. Dolphin3-Cyber-8B is also the only entry with throughput figures, 5 tok/s on CPU and 55 tok/s on an RTX 4060. ## Five different ways to strip a refusal The uncensoring column mixes methods that are not close to substitutes, and the README treats them as one field. DeepHat V2 went through supervised fine-tuning on 1.7 million offensive and defensive samples, described as coming from a USENIX Security 2024 workshop. BugTraceAI-CORE-Apex was tuned on HackerOne Hacktivity 2024-2025 plus a WAF evasion dataset. Cyber-Prime 1.1 is the outlier: it adds reinforcement learning and reward-guided post-training over 75K cybersecurity rows, broken down into NER repair around 6K examples, HTTP reasoning with chain of thought around 5K, email phishing around 5K and threat intel summarization around 2K, plus GHSA, KEV and ATT&CK data. Dolphin3-Cyber-8B takes a lighter path, a LoRA rank-16 applied on top of an already abliterated base. Cyber-Ornith-1.5-9B uses an obliteration variant. CYBER-FROST-3.8 was fine-tuned on a proprietary Blackfrost-AI corpus spanning recon, web app security, vulnerability research, malware analysis, cloud security and threat intel. Four of these methods are reproducible only by the model owner, since two of them depend on corpora the list does not publish. ## Offensive tooling and defensive analysis share one table The specialization row is where the list quietly mixes two different jobs. BugTraceAI-CORE-Ultra is labelled a tooling model and is described as generating Nuclei templates, CVE proof-of-concepts, exploit code and pentest scripts. Cyber-Prime 1.1 leans the other way, with training data split across phishing email work, HTTP attack reasoning, named entity recognition and threat intel summarization, and it exposes two operating modes: a direct mode for classification and a Think mode for chain-of-thought reasoning. CyberPal 2.0 is explicitly defensive, covering CTI, vulnerability analysis, detection and mitigation, SOC and incident response, AppSec and compliance work. Imperum-CybersecurityLLM v1.0 spans ten or more domains including SOC and SIEM operations, detection engineering, DFIR, malware analysis, threat intel, vulnerability management, cloud, Kubernetes and IAM, OT security, GRC and authorized pentesting. So a reader shopping for a triage assistant and a reader shopping for a template generator are picking from the same page, and the table format does not separate them. ## Licences vary more than the entries do Most entries carry Apache 2.0, including DeepHat V2, both BugTraceAI models, Cyber-Ornith-1.5-9B and Imperum-CybersecurityLLM v1.0. Three do not. CYBER-FROST-3.8 is under the Qwen Community License 1.0, Cyber-Prime 1.1 under the LFM Open License v1.0, and Dolphin3-Cyber-8B under Llama 3.1. A separate licence question belongs to the repository itself, which lists no licence, and its two top-level entries, README.md and offsec-capability-v3.png, include no LICENSE file. For an organisation planning to point internal traffic at one of these models, the practical consequence is that the licence travels with the weights, not with the catalogue, and that the catalogue carries no terms of its own. Anyone reading this as legal advice should instead read each licence text, but the structural point stands: selecting an entry from this list starts a licence review that a plain command line tool would not. ## Where the weights live, and the setup steps nobody wrote Each entry ends with a Download line pointing at a HuggingFace page, for example DeepHat V2 at https://huggingface.co/WhiteRabbitNeo and BugTraceAI-CORE-Ultra at https://huggingface.co/BugTraceAI/BugTraceAI-CORE-Ultra-27B-Q6. The README offers no install command, no serving configuration, no runtime invocation and no hint about which runtime fits which quantisation, so a first real use means choosing a runner for GGUF or safetensors files yourself and pulling the weights from the listed page. That is a deliberate gap for a curated list, but it pushes the actual cost onto the reader, who must also confirm whether a model with Tool Calling set to Yes needs a specific chat template or function-calling harness, and whether a model with Tool Calling set to No, such as Cyber-Prime 1.1, CyberPal 2.0 or Dolphin3-Cyber-8B, can be driven by an agent loop at all. The repository cannot answer either question, because it never describes a runtime. ## One benchmark score, one picture, and a September snapshot Only one entry reports evaluation numbers. Cyber-Prime 1.1 lists a CyberBench average of 0.592 F1/Acc, up from 0.501 in v1.0, with sub-scores of NER 0.499, Phishing 0.890 and HTTP Attack 0.628. The README gives no dataset description, no baseline models and no method, so the figure cannot be reproduced or compared with anything else on the page. The other entries carry no scores at all, which makes the list a catalog of claims rather than a benchmark result. The README also notes the data was gathered in Sep 2026, and the repository records no GitHub releases. The last push was on 2026-09-28. It is not archived, but a list whose entries change when model cards change has no release rhythm you can plan around, and the one image it ships is named offsec-capability-v3.png while the README refers to it with different alt text, a small sign that the asset and its caption drifted apart. ## Conclusion Adopt this repository as a shortlist and a hardware planning sheet, not as a dependency, and treat it as one if you can run a 2.6B model on a laptop or a 27B model on a 24 GB card. Skip it if you need documented setup steps, reproducible evaluation, or a single tool to call, because the README documents none of those and the repository itself carries no licence file. Verify three things first: the context length on the model card you actually pull, the licence that applies to your chosen entry, and whether your hardware meets the stated VRAM figure at the quantisation the table names. ## FAQ ### What is Offensive-Security-AI-Models and what does it contain? It is a curated list of open-weight uncensored models or models fine-tuned for cybersecurity tasks, aimed at authorized red team operations, penetration testing and security research. The repository holds only README.md and one image; the content is a set of comparison tables covering ten visible entries. ### Which model in the list is the easiest to run on a single consumer GPU? Several entries list roughly 6 GB of VRAM: DeepHat V2 at 7B or 32B with a 131K context, Cyber-Prime 1.1 at 2.6B in BF16, and Dolphin3-Cyber-8B at 8.03B in Q4_K_M. Cyber-Ornith-1.5-9B is listed at roughly 7 GB, and CYBER-FROST-3.8 is the outlier that needs multiple GPUs. ### Does the repository document how to install or serve these models? No. The README provides only Download links to the HuggingFace pages for each entry and no install command, runtime configuration or serving instructions. It states that the data was sourced from HuggingFace model cards and official publications as of Sep 2026. ### Which models in the list can use tools or agent workflows? DeepHat V2, both BugTraceAI models, CYBER-FROST-3.8, Cyber-Ornith-1.5-9B and Imperum-CybersecurityLLM v1.0 are listed with Tool Calling set to Yes. Cyber-Prime 1.1, CyberPal 2.0 and Dolphin3-Cyber-8B are listed with Tool Calling set to No. ### Do the listed models come with a benchmark result? Only Cyber-Prime 1.1 reports one, a CyberBench average of 0.592 F1/Acc against 0.501 in v1.0, with NER 0.499, Phishing 0.890 and HTTP Attack 0.628. The README gives no method or dataset detail for that score, and no other entry lists evaluation figures. ## Sources - [Issues](https://github.com/JoasASantos/Offensive-Security-AI-Models/issues) - [JoasASantos/Offensive-Security-AI-Models on GitHub](https://github.com/JoasASantos/Offensive-Security-AI-Models) - [README](https://github.com/JoasASantos/Offensive-Security-AI-Models/blob/main/README.md) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/joasasantos-offensive-security-ai-models