SecureCLI-Tuner implements a Zero-Trust Security Kernel that validates every generated command before execution.

User Query → Router → Generator (LLM) → CommandRisk Engine → Secure Wrapper → Execution
                                              ↓
                              ┌───────────────┼───────────────┐
                              ↓               ↓               ↓
                        Deterministic    Semantic       Policy
                        Guardrail        Guardrail      Guardrail
                              ↓               ↓               ↓
                         SigmaHQ         Intent ML       OWASP ASI

• Purpose: Hard blocks via regex patterns
• Coverage: 17+ zero-tolerance patterns
• LLMSVS: V2 (Model Operating Environment)
• Response Time: <1ms

• Purpose: Risk scoring
• Method: Command complexity + target sensitivity + MITRE mapping
• Threshold: Score >70 triggers block
• LLMSVS: V5 (Model Monitoring)

• Purpose: Intent-to-action alignment
• Method: ML classifier + obfuscation detection
• Coverage: Base64, hex, eval, nested substitution
• LLMSVS: V1 (Architectural Design)

1. User input is untrusted — May contain prompt injection
2. LLM output is untrusted — May hallucinate dangerous commands
3. Only validated commands pass — After 3-layer inspection

All blocked commands are logged with:

• Timestamp
• Original command
• Blocking guardrail
• ASI ID attribution
• MITRE ATT&CK technique
• Risk score

Logs are sent to W&B SecureCLI-Monitoring for real-time alerting.