squ11z1/Mythoseek

Mythoseek: A Cybersecurity-Specialized LLM

Mythoseek is a 10 billion parameter language model developed by Merlin Research as part of the KAON quantum-classical research program. Fine-tuned from DeepSeek V4 Pro-Qwen3.5 9B Distilled, it is specifically designed for cybersecurity applications, leveraging enterprise pentest reports and frontier model distillation traces.

Key Capabilities & Features

• Cybersecurity Specialization: Optimized for vulnerability research, penetration testing, OSINT, and CWE-pattern reasoning.
• Quantum-Classical Integration: Utilizes OTOC scrambling measurements from IBM Quantum hardware to calibrate Adaptive Entropy Regularization (AER) coefficients during GRPO training.
• Performance on CyberGym: Achieves 62% pass@4 on Level 0 (full scaffolding) of the CyberGym benchmark, a large-scale cybersecurity benchmark of real-world vulnerabilities.
• Training Methodology: Employs a multi-stage pipeline including SFT Distillation, GRPO/RL with verifiable rewards on cyber tasks, tool-use SFT, and CWE Grounding.

Intended Use Cases

• Vulnerability research and CVE analysis
• Assistance with penetration testing (OSINT, recon, XSS, SQLi)
• CWE classification and pattern recognition
• Security report generation
• Red team reasoning support

Mythoseek provides an open-weight alternative for advanced cyber AI, bringing capabilities previously found in closed-source models to the community.