Security

Guardrails

Problem

Users inject prompts ('ignore previous instructions'), share PII, or receive harmful outputs. Production LLM apps need defence in depth.

Pattern

Run an input guard before the model and an output guard before the response. Block, redact, or fall back to a safe response.

Implementation

guardrails.pypython

def input_guard(user_text: str) -> str:
    # PII redaction
    text = redact_pii(user_text)
    # Injection detection
    if detect_injection(text):
        raise BlockedError("Possible prompt injection")
    return text

def output_guard(model_output: str) -> str:
    if detect_secret(model_output):
        return redact_secrets(model_output)
    if detect_harmful(model_output):
        return SAFE_FALLBACK
    return model_output

@app.post("/api/chat")
def chat(req):
    safe_input = input_guard(req.message)
    raw = call_llm(safe_input, req.history)
    safe_output = output_guard(raw)
    audit_log(req.user_id, safe_input, safe_output)
    return safe_output

Trade-offs

Defends against injection and leakage
Adds latency (guard calls)
False positives degrade UX
Guard models themselves can be bypassed

Production checklist

Input + output guards
PII redaction
Secret detection
Audit log of blocked content
Red-team testing
Bypass monitoring