Data & Adversarial Risks

V2 console

This card lives in Policy Engine > LLM Gateway at web.quilr.ai/policy. Edits join the shared draft and take effect once you review and publish a revision.

Decide what happens when PII, PHI, financial data, secrets, prompt attacks or your own custom detections appear in a model request or response.

Sections

Section	What it holds	Empty state
Detection rules	Which data types are found, how many times, and the action.	Nothing is redacted or blocked; findings are still detected and reported.
Sensitivity profile	Detection sensitivity per category, subcategory actions, and scan scope for the whole request.	Engine defaults: every category detects at the engine's default sensitivity.

Applies on assistants, bedrock, chat, copilot, embeddings, rerank, responses, sdk_check, stt, text, tts and vertex.

Detection rule settings

Add rule opens the New data rule dialog.

Setting	Options	Default	Notes
Data types	Whole category, individual types, or custom detections	None	Choose types... lists every category. Several types on one line match any of them.
Findings threshold	at least N	1	Per line.
Action	Monitor, Partial redact, Redact, Block	Monitor	Block on any line stops the whole request.
Stage	Request, Response, Both	Request	Response redaction and blocking need the full response; streamed responses are scanned and tagged, not changed.
Tool boundary	Scan tool-call arguments	Off	Judges each tool call's arguments on their own. Only Monitor and Block apply.
Severity	Not set, Very low, Low, Medium, High, Critical, Very critical	Not set	Reported for dashboards, exports and alerts. Never changes the action.
More options	Priority, extra rules, metadata and content conditions, raw QuilrQL	-	Opens the full editor.

Add more detection lines to one rule with + Add line. Each line keeps its own types, threshold, action and stage.

Example

block_request_secretsrequest

runs on request·priority 900

Whendata foundis any ofAuth & Secrets

Then

Sensitive data action→block

Risk level→critical

Start new detections on monitor, review activity, then raise the action.

Scoping and precedence

• Applies to: Everyone, People, Smart group, Application, App tag, Requested model, Provider, API surface, Environment, Prompt complexity, Prompt text, Tool, Source network, or Except.... Every chip must match; values accept * and ?.
• Everyone is priority 500. Narrower scopes win over it.
• Highest priority wins per data type. At equal priority the more restrictive action wins.
• redact and partial-redact rewrite only the findings their own rule selected.

More examples, the full data type catalog and multi-rule configurations: LLM Gateway Policies.

Legacy app setting

Replaces the data risks, adversarial risks and precision detections parts of an app's Security Guardrails. Custom detections themselves are still defined per app under Custom Intents.