Without filter mode off
Sent to the model
Agent reply mock agent
Stored in the chat log
With filter mode redact
Sent to the model
Agent reply mock agent
Stored in the chat log
Summary of exposure
| Without filter | With filter |
|---|
Also in the pull request experimental · not deployed yet
- Tasks. In redact mode, string values in a task's JSON input
are redacted and the task runs. Block mode answers 422, naming the types.
Schema errors never quote a value.
{"name": "ClientError", "message": "Input blocked by guardrails", "details": {"category": "pii", "entities": ["PHONE"]}} - Block names the type. The refusal is a stream event with category
pii. Its message andentitiesname the types, never the values.{"type": "data-guardrail-violation", "data": {"category": "pii", "guardrailType": "input", "fallbackResponse": "Your message seems to contain sensitive personal data (payment card number), so it was not processed. Remove it and try again.", "entities": ["CREDIT_CARD"]}} - Detection. It uses kit v0.2.1, and also catches grouped or lowercase IBANs, a card followed by its expiry, and cards split by dots, tabs, line breaks or spaced hyphens. A phone with no + or spaces needs a word like “phone” before it and no order word between. Only some French, UK, German and 00 forms are caught.
- Only new input blocks: user messages after the last agent reply. Earlier turns and replies sent back are redacted silently. So is a stored message resent unchanged.
- Limits. A request with over 32,000 new characters is refused. Tool data and the system prompt are not checked. ≈1.6% of EAN-13 barcodes and ≈2.6% of 16-digit order numbers read as cards. An identifier split across two JSON values of a task input is missed.
What it finds
Use your own agent optional
- The left side sends the raw message. Use made-up numbers only.
- The key stays in this tab's memory and goes only to your app. Reload to forget it.
- Your agent keeps each conversation under its own retention setting.