philbert440/Qwen3.6-27B-Uncensored-Cyber
philbert440/Qwen3.6-27B-Uncensored-Cyber is a 27 billion parameter Qwen3.6-27B vision-language model, specifically modified by Rivet (philbert440) using the Heretic method. It is refusal-suppressed for cyber/offensive-security prompts, enabling responses to technical hacking, malware, and vulnerability queries while retaining guardrails against personal harm, doxxing, weapons, disinformation, and NSFW content. This model is optimized for authorized red-teaming, security research, CTFs, and education, providing real, working code and step-by-step instructions for offensive security tasks.
Loading preview...
Model Overview
philbert440/Qwen3.6-27B-Uncensored-Cyber is a specialized 27 billion parameter vision-language model based on Qwen3.6-27B, developed by Rivet (philbert440). This model has been uniquely modified using the Heretic method, which employs KL-optimized, per-layer continuous abliteration, to specifically target and suppress refusals for cyber/offensive-security-related prompts.
Unlike maximally open models, this version is deliberately targeted: its refusal objective was scored exclusively on a cyber-only prompt set. This allows it to provide technical offensive-security information, including real, working code for hacking, malware, and exploitation, while retaining guardrails against content related to personal harm, doxxing, weapons, disinformation, and NSFW topics. The model's vision, video, and MTP speculative-decoding head are fully preserved, ensuring no damage to general fluency or coherence on neutral long-form prompts.
Key Capabilities
- Offensive Security Focus: Provides real, working code and step-by-step instructions for hacking, reconnaissance, malware generation (e.g., file-encrypting ransomware), and physical security (e.g., lockpicking).
- Refusal Suppression: Overcomes base model refusals for technical offensive-security queries.
- Retained Guardrails: Deliberately refuses requests concerning doxxing, privacy invasion, weapons, disinformation, and explicit content.
- High Coherence: Maintains full-length, on-topic, and well-structured answers without degeneration on neutral prompts.
- Vision and MTP Preserved: Fully supports vision tensors and speculative decoding, functioning like the base Qwen3.6-27B VL model.
Good For
- Authorized Red-Teaming: Assisting in penetration testing and security assessments.
- Security Research: Exploring vulnerabilities and developing countermeasures.
- CTFs (Capture The Flag): Providing technical guidance for cybersecurity challenges.
- Education: Learning about offensive security techniques in a controlled environment.
This model is intended for use by consenting professionals and researchers, who are responsible for its ethical application.