Reference

What Omna hides

Every time your team sends something to an AI, it can carry secrets and personal data — keys in a config file, a client's email, an ID number in a spreadsheet. Omna finds them on your own computer and masks each one with a label before anything leaves.

3 min readSecrets and personal data

How it works

Each value Omna finds is masked — replaced by a bracketed label plus a number. The same value always gets the same label, so the AI can still follow the conversation, and your real values come back in the answer.

text
Jane Doe          →  [PERSON_1]
jane@acme.co      →  [EMAIL_1]
(415) 555-0132    →  [PHONE_1]
123-45-6789       →  [GOV_ID_1]
4111 1111 1111 1111  →  [FINANCIAL_1]
AKIA…EXAMPLE      →  [SECRET_AWS_KEY_1]

The AI receives only the masked labels. It never learns the real values.

Secrets — 240 kinds

Passwords, API keys, access tokens, private keys and database logins: AWS, Google Cloud, GitHub, Stripe, OpenAI, Anthropic, JWTs, SSH and private keys, database connection strings, and more.

Secrets are handled more strictly than personal data. Their real values are held only in memory, never written to disk, and put back only in the text you read — never in the commands an AI agent runs or the files it writes, so a poisoned web page can't trick the agent into sending your key somewhere. Secrets are always hidden; they can't be switched off.

KindMasked label
Cloud keys[SECRET_AWS_KEY_N]
Code-hosting tokens[SECRET_GITHUB_TOKEN_N]
API keys[SECRET_API_KEY_N]
Login tokens (JWT)[SECRET_JWT_N]
Passwords[SECRET_PASSWORD_N]
Private keys[SECRET_PRIVATE_KEY_N]
Database logins[SECRET_DATABASE_CONNECTION_STRING_N]
Other secrets[SECRET_GENERIC_SECRET_N]

Personal and client data

CategoryExamplesMasked label
NamesPeople in prompts and files[PERSON_N]
Contact detailsEmail addresses, phone numbers, postal addresses[EMAIL_N], [PHONE_N], [ADDRESS_N]
Government IDsSSNs, national IDs, passports, driver's licences, tax IDs — 50 ID formats with real checksum checks[GOV_ID_N]
Financial detailsCredit cards, bank accounts, IBANs, crypto addresses[FINANCIAL_N]
Health detailsPatient and medical record numbers, lab accession numbers, insurance IDs, diagnosis codes[HEALTH_N]
Work detailsEmployee IDs, salaries[EMPLOYMENT_N]
Digital identifiersIP addresses, device IDs, MAC addresses[DIGITAL_N]
Dates of birthBirth dates[BIRTH_N]

Government IDs include India's: Aadhaar, PAN, GSTIN, Voter ID, passport, driving licence, UPI IDs, vehicle numbers, ABHA, UAN and PIN codes. Names and addresses written as ordinary sentences are caught by smart masking, which is on by default. File names are hidden the same way as the file's contents.

For health data, Omna hides who the patient is, not what is wrong with them. Hidden: names, contact details, addresses, clinic and hospital names (shown as [ADDRESS_N]), and numbers with a label next to them ("patient ID", "Patient #", "MRN", "medical record number", "accession #", "NPI", "insurance ID", "diagnosis:"). Not hidden, so the AI can still help: illnesses, medicines and doses, lab results and procedures written in plain words, and the dates of visits and treatments.

Every category except secrets can be switched off: see Where masking stops.

Known limitations

Omna is built to miss as little as possible while leaving ordinary text alone, but no automatic system catches everything:

  • Unusual names — single names or rare spellings may occasionally be missed.
  • Initials — short initials like "J.S." are not flagged, to avoid false alarms.
  • Numbers without context — a long number is only treated as a card or ID when it looks like one.
Note: Omna is a privacy tool, not a compliance certification. If you're subject to HIPAA, GDPR or similar rules, check with your legal team how Omna fits your obligations.

What's next