Phishstral

A fine-tuned version of Mistral-7B-Instruct-v0.1 by Phishers of Men, optimized for security research and ethical hacking contexts.

Model Description

Phishstral is a 7B parameter language model fine-tuned with LoRA adapters on a curated dataset combining Python code, OpenHermes instruction data, and security-focused content. It is designed to assist security researchers, penetration testers, and CTF players with technical tasks where general-purpose models apply excessive filtering that impedes legitimate research workflows.

The model has reduced safety filtering compared to the base model in security research contexts, allowing it to discuss offensive security techniques, vulnerability classes, and related tooling more freely than standard instruction-tuned models.

Intended Use

Intended users:

• Security researchers
• Penetration testers
• CTF (Capture the Flag) participants
• Ethical hackers and red teamers
• Students learning offensive security

Intended use cases:

• Explaining vulnerability classes and exploitation techniques
• Writing and reviewing security-relevant Python code
• Assisting with CTF challenges
• Generating payloads and proof-of-concept code for authorized testing
• General-purpose instruction following with a security mindset

Training Details

Property	Value
Base model	mistralai/Mistral-7B-Instruct-v0.1
Fine-tuning method	LoRA (via Unsloth)
Training data	Python code, OpenHermes 2.5, security research data
Merged	Yes (full weights, not adapter-only)

Prompt Format

Phishstral uses the standard Mistral instruct format:

<s>[INST] {system_prompt}

{user_message} [/INST] {assistant_response} </s>[INST] {user_message} [/INST]

Limitations

• 7B parameters — capable but not on par with larger models for complex multi-step reasoning
• Training is ongoing; early release
• Like all LLMs, can hallucinate technical details — always verify output

About

Built by Phishers of Men — a security research group.