Pular para o conteúdo

PII detection

Atualizado em Ver como Markdown

AI Security for Apps (formerly Firewall for AI) can detect personally identifiable information (PII) in incoming LLM prompts. There are two approaches to PII detection, and you can use them together for layered protection:

  • Fuzzy detection (AI-powered) — AI Security for Apps uses an AI model to identify common PII types in the prompt content. This approach catches PII even when it appears in natural language or unexpected formats.
  • Exact detection (regex) — You write a WAF custom rule with a regular expression on the raw request body. This approach is ideal for organization-specific identifiers with a known, predictable format.

Fuzzy PII detection

When AI Security for Apps is enabled and a request arrives at a cf-llm labeled endpoint, it scans the prompt for PII and populates two fields:

The detection is based on Presidio ↗, a data protection and de-identification SDK. Refer to the cf.llm.prompt.pii_categories field reference for the full list of recognized categories.

Supported PII categories
Category Description
CREDIT_CARD Credit card number
CRYPTO Cryptocurrency wallet address
DATE_TIME Date or time expression
EMAIL_ADDRESS Email address
IBAN_CODE International bank account number
IP_ADDRESS IP address
NRP Nationality, religious, or political group
LOCATION Physical location or address
PERSON Person name
PHONE_NUMBER Phone number
MEDICAL_LICENSE Medical license number
URL URL
US_BANK_NUMBER US bank account number
US_DRIVER_LICENSE US driver license number
US_ITIN US Individual Taxpayer Identification Number
US_PASSPORT US passport number
US_SSN US Social Security Number
UK_NHS UK National Health Service number
UK_NINO UK National Insurance Number
ES_NIF Spanish tax identification number
ES_NIE Spanish foreigner identification number
IT_FISCAL_CODE Italian fiscal code
IT_DRIVER_LICENSE Italian driver license
IT_VAT_CODE Italian VAT code
IT_PASSPORT Italian passport number
IT_IDENTITY_CARD Italian identity card
PL_PESEL Polish national identification number
SG_NRIC_FIN Singapore National Registration Identity Card / Foreign Identification Number
SG_UEN Singapore Unique Entity Number
AU_ABN Australian Business Number
AU_ACN Australian Company Number
AU_TFN Australian Tax File Number
AU_MEDICARE Australian Medicare number
IN_PAN Indian Permanent Account Number
IN_AADHAAR Indian Aadhaar number
IN_VEHICLE_REGISTRATION Indian vehicle registration number
IN_VOTER Indian voter ID
IN_PASSPORT Indian passport number
FI_PERSONAL_IDENTITY_CODE Finnish personal identity code

Be specific to reduce false positives

The cf.llm.prompt.pii_detected field returns true when any PII category is detected — including broad categories like PERSON, DATE_TIME, and LOCATION that frequently appear in normal conversation. Blocking based on this field alone will produce a high false-positive rate for most applications.

Instead, build rules against cf.llm.prompt.pii_categories and list only the categories that matter for your use case. For example, a customer support chatbot may need to block credit card numbers and SSNs but can safely ignore person names and dates. Start with the narrowest set of categories, monitor matches in Security Analytics, and expand only as needed.

Example rules — fuzzy detection

Block any request containing PII

  • When incoming requests match:

    Field Operator Value
    LLM PII Detected equals True

    Expression when using the editor:
    (cf.llm.prompt.pii_detected)

  • Action: Block

Block only specific PII categories

  • When incoming requests match:

    Field Operator Value
    LLM PII Categories is in Credit Card

    Expression when using the editor:
    (any(cf.llm.prompt.pii_categories[*] in {"CREDIT_CARD"}))

  • Action: Block

Log email addresses but block credit cards and SSNs

Create two custom rules:

  1. A rule with action Block and the following expression:
    (any(cf.llm.prompt.pii_categories[*] in {"CREDIT_CARD" "US_SSN"}))

  2. A rule with action Log and the following expression:
    (any(cf.llm.prompt.pii_categories[*] in {"EMAIL_ADDRESS"}))

Exact PII detection (regex)

If you need to detect custom PII formats specific to your organization — such as internal employee IDs, patient record numbers, or proprietary account identifiers — you can create a WAF custom rule using a regex match on the raw body (http.request.body.raw field).

This approach complements fuzzy detection by covering formats the AI model does not natively recognize.

Example: Detect employee IDs

In the following example, an organization uses employee IDs in the format EMP- followed by exactly six digits (for example, EMP-482910).

Create a custom rule with the following configuration:

  • When incoming requests match:

    Field Operator Value
    Raw request body matches regex EMP-[0-9]{6}

    Expression when using the editor:
    (http.request.body.raw matches "EMP-[0-9]{6}")

  • Action: Block

  • With response type: Custom JSON

  • Response body: { "error": "Request blocked: employee ID detected in prompt." }

Scope to a specific endpoint

To limit this rule to only your LLM endpoint, combine it with a path condition:

Field Operator Value Logic
URI Path equals /api/chat And
Raw request body matches regex EMP-[0-9]{6}

Expression when using the editor:
(http.request.uri.path eq "/api/chat" and http.request.body.raw matches "EMP-[0-9]{6}")

More regex examples

Custom PII type Example format Regex pattern
Employee ID EMP-482910 EMP-[0-9]{6}
Patient record number PAT/2024/00391 PAT/[0-9]{4}/[0-9]{5}
Internal account ID ACCT-XX-99999 ACCT-[A-Z]{2}-[0-9]{5}
Custom API key prefix sk_live_abc123... sk_live_[a-zA-Z0-9]{20,}

Considerations for regex rules

  • Cloudflare Plan requirement. Regex operators (matches and ~) require a Business or Enterprise plan.
  • Body size limit. The http.request.body.raw field inspects a limited portion of the request body. The exact limit varies by plan.
  • JSON payloads. The raw body includes the full JSON structure. Your regex should account for the fact that the prompt text is nested inside a JSON string.
  • Performance. Complex regex patterns can impact rule evaluation time. Keep patterns as specific as possible.

Combine both approaches

You can use fuzzy and exact detection together for layered protection:

(cf.llm.prompt.pii_detected or http.request.body.raw matches "EMP-[0-9]{6}")

This rule blocks requests where either the AI model detects any built-in PII category or the regex matches your custom identifier format.