README.md

---
title: OpenMythos
emoji: 🛡️
colorFrom: gray
colorTo: indigo
sdk: gradio
sdk_version: 6.18.0
python_version: '3.13'
app_file: app.py
pinned: true
short_description: An Open Source Cyber Security Agent
license: apache-2.0
---

# openMythos 🌌

**Paste your codebase. Our AI security agent audits the repository** — a multi-level vulnerability analysis, a visual dependency risk path, a declared threat level — then generates an instant, verifiable hotfix patch before threat actors can exploit it.

Built during the **Hugging Face Small Gradio Hackathon**, openMythos democratizes cutting-edge security auditing. It bridges an immersive retro terminal interface with the elite agentic reasoning and long-context preservation architecture of a fine-tuned dense model.

> ⚠️ **Proactive Defense.** This platform is engineered for defensive security intelligence. It aims to discover flaws, memory leaks, security configurations, and input bugs instantly, empowering software engineering teams to deploy hotfixes long before a threat vector is weaponized.

---

## ▶️ See it in action

- **Demo video:** TODO — Watch the Social Media Demo Video & Technical Explainer Post
- **Social post:** TODO — Paste your launch post link here

---

## Why it's worth a look

- 🧠 **Deep Agentic Reasoning, Not a Basic RegEx Scanner.** Powered by a specialized Qwen3.6-27B foundation architecture, openMythos maps complex variable trails and dependency structures across entire software repositories during a single security sweep using its native long-context window.

- 🎨 **Immersive Retro UI.** No default Gradio look: a distraction-free retro terminal architecture optimized for low-latency code-auditing loops.

- 🔌 **100% Local & Privacy-First.** Designed as a fully open-source alternative to proprietary security intelligence layers (like Claude's Mythos model). It can be run entirely locally, requiring zero internet connectivity or external dependencies to operate.

---

## How it works

A multi-stage engineering pipeline built around aggregated, industry-standard security sources:

| Stage | Role | Source Data / Methodology |
|:-----:|------|---------------------------|
| **1** | **Data Prep & Aggregation** | Incident reports, GitHub Advisory, VulnHub, and papers. Rigorously trained on BigVul-Filtered and Arvix-Filtered sets. |
| **2** | **Initial Fine-Tuning (SFT)** | Supervised Fine-Tuning on cybersecurity tasks. Qwen3.6-27B Base (Up to 262k+ token context window). |
| **3** | **Reinforcement Learning (RLVR)** | Verifiable Reward via vulnerable vs. fixed repo branches. Verified by a separate evaluation model checking fixes. |
| **4** | **Rigorous Evaluation** | Benchmarked against CyberGYM and SWE Bench Verified. Evaluates historical vulnerabilities and code generation. |

The entire pipeline leverages highly specialized weights to ensure an elite vulnerability discovery rate. No massive API dependencies anywhere: a clever chain of targeted engineering (**prepare → SFT → RLVR → verify**) delivers the whole security suite.

```
Raw Codebase Input
   └─▶ Stage 1: Data Prep  ─ BigVul & arXiv research paper data curation
        └─▶ Stage 2: SFT Train  ─ Supervised fine-tuning on targeted cybersecurity tasks
             └─▶ Stage 3: RLVR Refinement  ─ Reinforcement Learning via Verifiable Rewards (Vulnerable vs Fixed Code)
                  + CyberGYM & SWE Bench verification models
                  + Retro Terminal UI output
                  → Instantly remediated source-code patch
```

---

## Tech

- **Frontend:** This Gradio 6 Space using an immersive terminal configuration.
- **Base Architecture Alternative Options:** While utilizing Qwen3.6-27B, the training framework also supports Devstral-Small-2-24B, Magistral-Small, gemma-4-12B-it, and gpt-oss-20b.
- **Data Integrations:** Hardwired to ingest top-tier vulnerability streams like BigVul-Filtered and ArvixImport-Filtered-Final.

---

## Run it locally

```bash
# Clone the repository and initialize the security agent interface locally
python app.py
```

---

## 🤝 Project Contributors & Ecosystem Credits

Developed with ❤️ during the **Hugging Face Small Gradio Hackathon** by:

- **KingNish** – [HuggingFace Profile](https://huggingface.co/KingNish)
- **Himanshu** – [HuggingFace Profile](https://huggingface.co/Himanshu)

---

## 📜 Citations & Academic Attributions

```bibtex
@misc{openmythos2026,
    title  = {openMythos: Defensive Security Code-Auditing Agent Interface via Qwen3.6 Context Preservation},
    author = {KingNish and Himanshu},
    year   = {2026},
    howpublished = {Hugging Face Small Gradio Hackathon Project Suite}
}

@misc{qwen3.6-27b,
    title  = {{Qwen3.6-27B}: Flagship-Level Coding in a {27B} Dense Model},
    author = {{Qwen Team}},
    month  = {April},
    year   = {2026},
    url    = {https://qwen.ai/blog?id=qwen3.6-27b}
}