OpenMythos is a cybersecurity-focused LLM
It's trained specifically on:
- ArXiv cs.CR (Cryptography & Security) research papers
- Real CVE vulnerability data with detailed descriptions
What did we build?
- Full SFT + RLVR training pipeline on H100 GPUs (via Modal)
- Custom datasets: filtered ArXiv papers + CVE vulnerability data
- Deployed on Hugging Face Spaces
What was challenging?
- Scraping and filtering high-quality dataset
- Getting RLVR to actually improve security reasoning.
- Running stable distributed training on Modal with DeepSpeed
- Making the model actually useful for real vuln analysis, not just theory
What are we proud of?
- Model genuinely understands CVEs and can reason about attack surfaces
- Clean SFT → RLVR pipeline that others can replicate
- 9,000+ users in the first week with zero marketing budget
- Benchmark results on CyberGym and cybench
How to test it?
Hugging Face Space: https://huggingface.co/spaces/build-small-hackathon/OpenMythos
Try prompts like:
- "Analyze this code for security vulnerabilities: [paste your code]"
- "Review this API endpoint for security issues: [paste code]"
It works best for:
✅ Code security review
✅ CVE explanation & mitigation
✅ Attack surface analysis
✅ Security research questions
- 4 devlogs
- 50h