Finding and Masking PII

Finding and Masking PII

Step 1 of 9

raw PII3

A support message arrives: a name, a card number and an email, plus the actual question — why was I charged twice?

The Idea

Personal data leaks through three doors: the prompt, the reply and the logs. Detection finds each entity with a type, a confidence score and its position in the text. Masking swaps it for a typed placeholder such as {NAME}, so the model can still answer the question. Then scan the output, and mask logs at ingestion, because a log may keep the original.

Real-World Example

In a sealed-bid tender, the judges score bids with each company's name replaced by a number. The envelope matching numbers to names sits in a separate safe that only one person can open.

The Tradeoff

Detection is probabilistic: a low threshold masks harmless numbers, a high one misses real ones. On AWS, Amazon Comprehend detects PII, Amazon Bedrock Guardrails mask prompts and replies, CloudWatch Logs data protection masks log lines, and Amazon Macie finds PII already sitting in S3.

Your turn

Put the steps in the right order.

1. 1Scan the model's reply before the user sees it↑↓
2. 2Detect PII entities in the incoming message↑↓
3. 3Send the masked prompt to the model↑↓
4. 4Replace each entity with a typed placeholder↑↓

Check order

Mini quiz

1 / 3

Your gateway masks PII before the model sees it. Where can raw values still end up?

• AIn the model's weights
• BIn a request log that stored the original message
• CInside the placeholder text
• DNowhere, masking the input is enough