Sentinel guardrails: secrets, PII, and toxicity

SUPERWISE® Sentinel guardrails are the checks a Sentinel gateway runs on every inbound prompt before it reaches the model provider: secret and credential detection, PII redaction, rudeness and toxicity, language check, image resolution, and image sharpness. Most check the prompt's text, and two check the images attached to it. On a match, a guardrail either replaces the matched content with a placeholder and forwards the request, or blocks the request so it never reaches the model.

The checks run where the gateway runs, on the machine or in your network, so matched content is replaced before the request leaves it. For the general idea of one control point that every AI request passes through, see AI enforcement at runtime on superwise.ai.

ℹ️
Guardrails run on inbound prompts. Most let you choose whether to remediate (replace matched content with your placeholder text and forward the request) or block the request outright; a few, such as language check, image resolution, and image sharpness, can only block. See Manage guardrails.

Which guardrails are available?

Secret & credential detection: catch API keys, cloud credentials, and private tokens before they reach the model.
PII redaction: catch emails, phone numbers, government and financial IDs, and other personal identifiers.
Rudeness & toxicity: catch profanity, hostile language, and harassment.
Language check: block prompts that aren't in your allowed languages.
Image resolution: block prompts whose attached images are above or below a resolution limit.
Image sharpness: block prompts whose attached images are too blurry.

Each Sentinel runs its own guardrail policy. See Manage guardrails to add, remove, configure, and order guardrails on a specific gateway.

Get Sentinel: Sentinel on superwise.ai.

Which guardrails are available?

See also