Introduction to Arcjet Sensitive Info Protection

Arcjet sensitive information detection lets you detect and block sensitive data in request bodies before it enters your application. Use it to prevent clients from sending personally identifiable information (PII) and other data you do not want to handle.

What is Arcjet?

Arcjet is the runtime security platform that ships with your code. Enforce budgets, stop prompt injection, detect bots, and protect personal information with Arcjet's AI security building blocks.

Sensitive information detection is Arcjet’s AI Data Loss Prevention layer - stopping PII from leaking into AI model context, training data, or third-party tool calls made by agents. When a user pastes a credit card number or email address into a chat prompt, Arcjet catches it before it reaches your AI provider. All detection runs entirely locally in your own environment. No request body data ever leaves your infrastructure. To apply the same detection inside agent tool handlers and pipelines, see Arcjet Guards.

Detection runs through a pluggable backend. The built-in WebAssembly engine ships with the SDK and detects email addresses, card numbers, IP addresses, and phone numbers. For broader coverage, the optional @arcjet/sensitive-info-rampart package runs an on-device named-entity-recognition (NER) model that also detects names, addresses, and government / financial identifiers — all still local to your environment, with the model weights bundled so nothing is fetched at runtime. Either backend can be extended with custom patterns you define.

Sensitive info detection works on the request body. To detect specific information in headers or other request fields, use Arcjet filters. For AI endpoints, pair sensitive info detection with prompt injection detection to also block hostile instructions and jailbreak attempts before they reach your model.

When to use Arcjet sensitive info detection

Use sensitive information detection when you want to:

• Prevent PII from leaking into AI model context — stop users from accidentally (or deliberately) pasting card numbers, email addresses, or other regulated data into prompts sent to your AI provider.
• Block sensitive data from agent tool calls — prevent PII from appearing in tool arguments or responses that may be logged or forwarded to third-party services.
• Reduce your compliance burden by avoiding storage of PII and other regulated data.
• Prevent users from pasting card numbers or other sensitive data into generic text fields (for example, support forms, feedback forms, or chat boxes).
• Enforce rules that certain endpoints must not receive PII at all.

For email addresses specifically, you can also use Arcjet email validation to validate and verify addresses before creating accounts or sending email.

SDK-only rule

Sensitive information detection inspects the request body locally inside the SDK and is not available as a remote rule. Configure it in code with the SDK.

How Arcjet sensitive info detection works

Arcjet analyzes the body of each request locally in your application environment:

1. The request body is scanned by the configured detection backend.
2. The backend identifies sensitive entities — built-in structured types, and, with the Rampart backend, names, addresses, and other identifiers.
3. Any custom detection you configure (a detect function or a Rampart recognizer) also runs.
4. If sensitive information is found, Arcjet returns a decision so your application can block, sanitize, or otherwise handle the request.

All of this logic runs locally and in-process. The raw request body is never sent to Arcjet.

Detected entities

Which entities are available depends on the detection backend. The built-in engine detects four structured types; the on-device Rampart backend adds names, addresses, and government / financial identifiers:

Entity type	Built-in engine (default)	Rampart backend
EMAIL	✅	✅ Model + recognizer
PHONE_NUMBER	✅	✅ Model + recognizer
IP_ADDRESS	✅	✅ Recognizer
CREDIT_CARD_NUMBER	✅	✅ Recognizer
URL	—	✅ Model + recognizer
SSN	—	✅ Recognizer
GIVEN_NAME	—	✅ Model
SURNAME	—	✅ Model
TAX_ID	—	✅ Model
BANK_ACCOUNT	—	✅ Model
ROUTING_NUMBER	—	✅ Model
GOVERNMENT_ID	—	✅ Model
PASSPORT	—	✅ Model
DRIVERS_LICENSE	—	✅ Model
BUILDING_NUMBER	—	✅ Model
STREET_NAME	—	✅ Model
SECONDARY_ADDRESS	—	✅ Model
CITY	—	✅ Model
STATE	—	✅ Model
ZIP_CODE	—	✅ Model
Custom	✅ detect callback	✅ recognizers option

For the Rampart backend, Model means the type is detected by the on-device NER model, and Recognizer means it is detected by a deterministic, validated pattern (mirroring Rampart’s deterministic redaction layer). Where the model and a recognizer overlap on the same text, the recognizer wins.

The Rampart model is compact (a ~14.7 MB, 4-bit quantized artifact) and fast — around 6.6 ms median inference on Node.js — so it runs inline on each request, and recalls ~98% of private terms across the seven Latin-script languages it supports. See the reference and the model card for the full accuracy and latency breakdown.

You can combine either backend with your own custom detectors to cover additional data types specific to your application.

Privacy and reporting

Sensitive info detection runs entirely locally. Arcjet reports only the decision (for example, whether sensitive data was detected) to your dashboard; the actual request body content is not sent to the Arcjet Cloud API.

This lets you enforce PII detection and compliance controls while keeping sensitive data within your own environment.

Pricing

See the pricing page for details.

Discussion