TECHNIQUE

Input guards

Guardrails & Safety

1APPLICATIONS
2OBSERVED OPERATORS
01

State of Practice

CROSS-VALIDATED — 5 OPERATORS

Input guards are implemented as concrete pre-processing controls at gateways, APIs, RAG query paths, and privacy/data-enforcement boundaries; the pool shows different guard targets rather than one shared pattern.

Observed Practices

Filter or handle unsafe prompt/query inputs before downstream AI processing, including repeated-token prompts, prompt injection, jailbreaks, content safety, PII redaction, or RAG retrieval.

3 of 5 operators
OpenAIUberTraceIQ

Put guard and policy enforcement in mediation layers before requests reach downstream tools, services, model providers, or data systems.

3 of 5 operators
GrabUberMeta

Use privacy/data-policy guards to constrain ingestion, processing, access, and training-data use.

1 of 5 operators
Meta

Require review and checklist gates before new AI Gateway use cases are onboarded.

1 of 5 operators
Grab

Where Operators Diverge

Operators differ on where input guards sit in the system path.

APPROACH 01

Gateway/API mediation before model, provider, tool, or service access.

GrabUberOpenAI

APPROACH 02

RAG online query path before retrieval.

TraceIQ

APPROACH 03

Privacy-aware infrastructure at data ingestion, processing, access, lineage, and training-data boundaries.

Meta

Operators differ on what the guard is checking for.

APPROACH 01

Prompt-pattern filtering for repeated single-token prompts.

OpenAI

APPROACH 02

Prompt injection, jailbreaks, content safety, PII redaction, tool access checks, and sensitive-data redaction.

Uber

APPROACH 03

Access authorization, path-based provider/feature authorization, authentication, and rate limiting.

Grab

APPROACH 04

Privacy constraints and permitted-purpose checks on data assets before processing or training use.

Meta

APPROACH 05

Input guardrails before retrieval in a RAG system; the pool does not specify the exact checks.

TraceIQ

Watch Items

Single-token repetition filters did not cover all repeated-token divergence paths; multi-token repeats could still elicit model divergence and training-data extraction.

Repeated-token divergence could be used to bypass prompt guardrails and produce hallucinatory responses.

GPT-4 repeated-phrase behavior was non-deterministic in reported tests, and some repeat requests timed out after ten minutes.

02

Implementation Menu

CURATED DEFAULTS
NameKindMaturity
Prompt-injection classifier gatepatternestablished
Llama Guardlibraryestablished
Presidiolibraryestablished
03

Observed in Production

1 APP