Qwen3-1.7b_for_cybersec

Model Summary

Qwen3-1.7b_for_cybersec is a specialized, fine-tuned language model designed for AI Red Teaming and Offensive Security research. Built on the Qwen3-1.7B architecture, this model is optimized to assist cybersecurity professionals in auditing AI deployments and identifying vulnerabilities within LLM safety layers.

Unlike standard models with strict safety filters that may hinder security testing, this model is designed to be highly compliant with research-oriented prompts. It provides direct, technical assistance in simulating advanced threat vectors, including Prompt Injection, Jailbreaking, and Data Poisoning, specifically aligned with the newest AI threat vectors.

Intended Use

• AI Red Teaming: Simulating adversary behavior to test the robustness of AI systems.
• Vulnerability Assessment: Identifying weaknesses in system prompts and guardrails.
• Educational Purposes: Providing technical insights into LLM-specific attack vectors for authorized security audits.

Out-of-Scope Use

This model is NOT intended for any malicious activity or unauthorized access to systems. Use is strictly limited to legal AI audits where explicit permission has been granted.

Technical details & Fine-Tuning

The model underwent Supervised Fine-Tuning (SFT) to ensure it remains strict in its technical output and helpful during offensive security scenarios.

• Architecture: Qwen3-1.7B (Base)
• Method: QLoRA (Quantized Low-Rank Adaptation)
• Dataset: Custom-curated SFT/DPO dataset focused on modern AI threat vectors and security write-ups.

Fine-Tuning Hyperparameters

• Bits: 4-bit quantization
• LoRA Config: r=16, alpha=32
• Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Epochs: 3
• Dataset Size: 64 high-quality security examples

Dataset Link

This dataset was created by me and used for the fine tuning.

[Link to SFT dataset]

Risks and Ethical Considerations

This model is partially uncensored to allow researchers to test the absolute boundaries of LLM safety. Users must adhere to ethical hacking standards. The model provides brief, high-density technical answers and will not offer moralizing refusals when asked to generate attack-oriented content for testing purposes.

About the Author

Antoni Błoch – Cybersecurity student at AGH University of Krakow with a focus on LLM vulnerabilities and AI Red Teaming.