BYOLLM, Prompt Security & Auto-Triage

Bring Your Own LLM (BYOLLM) allows developers and SOC operators to plug in local or cloud AI models to power automated alert triage, investigation, and reporting.

---

Supported LLM Providers

LureGuard leverages opencode’s native provider ecosystem:

• Default: opencode/big-pickle
• Local Models: Ollama (Llama 3 8B, DeepSeek R1)
• Cloud Providers: OpenAI (GPT-4o, GPT-4o-mini), Anthropic (Claude 3.5 Sonnet, Claude 3.5 Haiku)
• OpenAI-Compatible APIs: Groq, OpenRouter, vLLM

---

Automated Alert Watcher (alert_watcher)

LureGuard includes an automated background worker (alert_watcher) that bridges incoming high-severity alerts to the AI analyst:

• Trigger Condition: Fires automatically whenever a Wazuh alert reaches rule level $\ge 12$ (configurable via AUTO_TRIAGE_LEVEL in .env).
• Requirements: opencode CLI must be available in system PATH.
• Execution: Spawns a headless session: opencode run "Read skills/triage.md — triage high severity alert".

---

Prompt Injection Defense Architecture

Attacker-controlled strings (usernames, HTTP user-agents, SSH banner payloads) are sanitized and wrapped inside strict boundary tags before submission to the LLM:

<<<ATTACKER_INPUT>>>
[Sanitized log payload harvested from auth.log / syslog]
<<<END>>>

System prompts instruct the AI model to treat data enclosed within <<<ATTACKER_INPUT>>> strictly as untrusted data, preventing prompt injection attacks from manipulating system commands or analyst verdicts.

Deterministic Security Policy Guarantee

LLM outputs serve strictly as EVIDENCE CONSUMERS and NEVER alter network policy or apply iptables rules automatically. All defensive actions require human confirmation or explicit rule enforcement.