Guardrails Reference

All guardrails surfaced by the inspection APIs.

The table below documents every guardrail surfaced by input-check and output-check, the engine class behind it, and the phase(s) in which it operates.

Guardrail	Phase	Engine	Description
LLM	Input + Output	LLM-as-judge	Scores prompt-injection likelihood 0–1 using a model-based judge.
Keyword	Input + Output	Rule-based	Exact keyword/phrase blocklist match.
Regex	Input + Output	Rule-based	Regex pattern match against operator-configured patterns.
PII Detection	Input + Output	NER model	Detects names, emails, phone numbers, NRIC/passport, credit cards, etc.
Vector	Input	Semantic similarity over TVDB	Similarity search against the proprietary Threat Vector Database. Catches paraphrased injection.
Content Moderation	Input + Output	Purpose-built classifier	Classifies harmful content (violence, hate speech, unethical, etc.).
System Prompt Protection	Output	In-line detection	Detects when the LLM is leaking its system prompt in its response.

When each guardrail fires

injection_detected is true in the API response if any active guardrail flags the input or output. Use the per-guardrail block within checks to identify which layer produced the decision and tune accordingly.

Configuring sensitivity

Guardrail	Tunable
LLM	Threshold (per security profile)
Vector	Sensitivity (Low / Medium / High)
Content Moderation	Category-level enable/disable
System Prompt Protection	Automatic in Forwarding mode (no application-side configuration required)
Keyword / Regex	Operator-defined rule sets
PII	Per-category toggles (NRIC/FIN, Phone, Email, Person, Credit Card, etc.)