Built-in detectors

PromptGlass ships with built-in detectors that classify the content of prompts, pastes, file uploads and (optionally) responses. Each detector emits a data class such as pci.card or secret.aws_key, a count, a confidence, and the character spans it matched. Policies act on data classes, not on individual detectors, so you can tune detection without rewriting policy.

Detection pipeline

Detectors run in tiers so the gateway stays fast:

1. Pattern + validator: regular expressions with checksums or structure checks (Luhn, IBAN mod-97, SIN/SSN rules, key prefixes). Sub-millisecond.
2. Dictionary and keyword proximity: medical terms, project code names, deal terms, near identifiers.
3. Exact data match (EDM): hashed lookups against your own records. See exact data match.
4. ML classifiers: named-entity recognition for people, organizations and addresses; a source-code classifier; document-type classifiers (contract, résumé, financial statement). Runs only when a rule needs it, or asynchronously for observe-only sensors.

Data classes

Data class	What it finds	Method
pii.email	Email addresses	Pattern
pii.phone	Phone numbers (E.164 and national formats)	Pattern + libphonenumber validation
pii.person_name	Personal names in context	ML NER
pii.address	Postal addresses	ML NER + pattern
pii.national_id.us_ssn	US Social Security numbers	Pattern + area/group rules + context
pii.national_id.ca_sin	Canadian Social Insurance Numbers	Pattern + Luhn
pii.national_id.uk_nino	UK National Insurance numbers	Pattern + prefix rules
pii.national_id.eu	Selected EU national IDs and tax IDs	Pattern + checksum
pii.dob	Dates of birth near identity context	Pattern + proximity
phi.medical_record	MRNs, health plan IDs	Pattern + context dictionary
phi.clinical	Diagnoses, medications, procedures (ICD-10, SNOMED-like terms) near a person	Dictionary + NER proximity
pci.card	Payment card numbers	Pattern + Luhn + IIN ranges
pci.track_data	Magnetic stripe track data	Pattern
financial.iban / financial.bank_account	IBAN, US routing + account, CA transit	Pattern + mod-97 / checksum
secret.aws_key	AWS access key IDs and secret keys	Prefix + entropy
secret.azure / secret.gcp	Azure connection strings, SAS tokens, GCP service-account keys	Pattern + structure
secret.api_key	OpenAI, Anthropic, GitHub, Slack, Stripe and other vendor tokens	Prefix patterns
secret.private_key	PEM/SSH private keys	Pattern
secret.password	Passwords in connection strings and config	Pattern + context
secret.jwt	JSON Web Tokens	Structure
code.source	Source code (with language)	ML classifier
code.proprietary	Code matching your repositories’ fingerprints	Fingerprint EDM
business.financials	Non-public financial results, forecasts	Keyword proximity + classifier
business.mna	M&A and deal terms, code names	Dictionary (customizable)
business.legal_privileged	Privileged & confidential legal text	Keyword + classifier
doc.label.*	Microsoft Purview / Google sensitivity labels on uploaded files	File metadata

Confidence and thresholds

Each match has a confidence of low, medium or high. Rules can require a minimum confidence and a minimum count:

match:
  data_class: [pii.national_id.*]
  min_confidence: medium
  min_count: 1

Testing a detector

promptglass policy test --text "Customer SIN 046 454 286, card 4111-1111-1111-1111"

pii.national_id.ca_sin   high   1   [13..24]
pci.card                 high   1   [31..50]
→ rule "redact-pci-and-national-ids" → action: redact

Tuning

• Allowlists: suppress known test values (4111 1111 1111 1111), your support phone number, public email domains.
• Context words: require nearby context for noisy classes (pii.dob needs “born”, “DOB” and so on).
• Per-app scope: turn on code detection only for AI coding tools, or skip it in approved internal assistants.

For organization-specific detectors, see custom detectors.