v0.12.5

NanoMind Semantic Analysis on All Scan Commands

• NanoMind now runs on every scan command: secure, check, scan-soul, secure-openclaw, secure-nemoclaw
• 7 new AST analyzers (prompt + code): jailbreak susceptibility, capability creep, command injection, unsafe deserialization, path traversal, injection resistance, authority confusion
• Auto-downloads model from HuggingFace on first use (~5.5MB, SHA-256 verified)
• scan-soul --deep uses NanoMind locally (free) before LLM fallback (paid)
• TME v4 model: 93.89% eval accuracy, all 9 attack classes including new credential_abuse
• Claude LLM chief data scientist reviewed all training labels
• Training pipeline: make pipeline in nanomind-training repo

Model

• Published to HuggingFace: opena2a/nanomind-security-classifier v0.4.0
• 1800 balanced training samples, 9 classes, 4 content types