majiayu000 ai-safety-guardrails Skill
This skill protects the system from its own AI.
1. Input Guardrails (Defense)
- Prompt Injection: "Ignore previous instructions".
- Defense: Delimiters (XML tags), "Sandwich Defense" (System Prompt + User Input + System Reminder).
- Jailbreaks: "Do this in 'DAN' mode".
- Defense: Pattern matching for known jailbreak signatures.
- PII Scrubbing: Regex scan input for SSN, Credit Cards, Emails before sending to LLM.
2. Output Guardrails (Verification)
- Hallucination Check: "Self-Consistency" (Ask 3 times, take majority).
- Tone Policing: Sentiment analysis on output. (Block Toxic/Aggressive responses).
- Format Validation: Ensure JSON is valid JSON.
3. Libraries & Tools
- NeMo Guardrails (NVIDIA)
- Guardrails AI (Python)
- Rebertha (PII)
4. System Design
- Human in the Loop (HITL): For high-stakes actions (Transfer Money), AI proposes, Human approves.
- Least Privilege: The Agent's API Token should NOT have admin access.