Security overview
How Obscate handles your data
For security and compliance reviewers. Everything here describes what exists today, unless it is marked planned. Last updated 19 November 2026 (hosting details added at launch).
Data handling
| Input text / JSON | Processed in memory for the duration of the request. Not written to disk, logs, or a database. |
|---|---|
| Detected values | Not stored or logged. |
| Restore map | Returned to you, never kept by Obscate. With the fpe strategy there is no map at all (token_map: {}). |
| Server logs | One line per request: timestamp, request id, method, path, status, latency — e.g. INFO rid=437516c03a00 POST /v1/redact 200 4ms. |
| Errors | Never echo the request body. Unexpected errors log only the exception type; tracebacks are never printed. |
| Account data | Email, plan, usage counters, the site page you signed up from (path only, e.g. /blog/ff1-explained), and an HMAC of each API key. Raw keys are shown once and never stored. |
These guarantees are enforced by automated privacy tests on both success and error paths.
What the hosted API can see: plaintext, in memory, while it processes your request. “Data never leaves your perimeter” is true only for the self-hosted edition (planned — design-partner pilots).
Cryptography
- NIST SP 800-38G FF1 with AES-256, tested against all nine official NIST FF1 sample vectors.
- FF1, not FF3-1: NIST’s February 2025 draft (SP 800-38G Rev. 1, 2nd public draft) drops FF3 after a published attack that also affects FF3-1.
- Key derivation: one 256-bit customer key → HKDF-SHA256 → separate FF1 and AES-SIV subkeys.
- Format- and checksum-preserving: cards stay Luhn-valid, Aadhaar Verhoeff-valid, IBANs pass mod-97, GSTINs keep a valid check character; PAN, SSN, phone, email (TLD kept), IPv4, MAC and AWS keys keep their format.
- Names, places, short values become authenticated AES-SIV tokens (
[PERSON:…]) — also map-free; restore never “decrypts” text that isn’t a genuine token. - Self-check: every
fperedaction verifies that restore reproduces the exact input before returning. - Deterministic: the same value gives the same token under the same key, so joins still work on redacted data.
Access control and abuse limits
- API keys are authenticated and requests rate-limited before the request body is processed.
- Per-key rate limits by plan (60 / 300 / 600 requests a minute) with
x-ratelimit-*headers; the Free quota is a hard cap. - 1 MB maximum request body. Up to 5 active keys per account, with rotation and revocation.
- Encryption keys come from the environment, never code; in production the server refuses to start without them.
- The public website demo uses separate, throwaway demo keys and is limited to 10 requests a minute per IP.
Hosting and data residency
- Processing: the hosted API runs on DigitalOcean App Platform in Bangalore, India (region
blr1). Request content is processed in memory there and never written to disk. - Account data (email, plan, usage counters, HMAC-hashed keys) is stored in managed Postgres in Bangalore: encrypted at rest, daily backups, 7-day point-in-time recovery, and accepting connections only from the app.
- Network path: traffic enters through Cloudflare’s edge network, which App Platform uses as its built-in CDN, then reaches the app in Bangalore. In our test from India, requests entered at Cloudflare’s Chennai location.
- Self-hosted (planned) keeps processing in your own region and network.
Service providers
| Provider | Purpose | Region | Data it handles |
|---|---|---|---|
| DigitalOcean | Hosting and managed Postgres | Bangalore, India | Request traffic in memory; account data at rest |
| Cloudflare | Edge network in front of the API | Nearest edge location | Request traffic in transit |
| Resend | Verification emails | Tokyo, Japan | The recipient’s account email address only. Open and click tracking disabled. |
| Sentry | Error reports | EU (Germany) | Exception type and stack frames only. Messages, request data and variables are stripped (verified with a real test event). |
Verification emails are sent from hello@send.obscate.com, with replies going to contact@obscate.com.
Regulation
Obscate is designed to help you meet DPDP Act obligations such as data minimisation and purpose limitation, and the equivalent GDPR principles. It is not a certification, and using it does not by itself make a system compliant.
Accuracy
On a synthetic 200-document benchmark generated by our test harness (python -m eval.benchmark; methodology): every structured or checksum ID type (card, Aadhaar, GSTIN, IBAN, PAN, SSN, email, phone, IPv4, MAC) scores F1 1.00; PERSON 0.80; LOCATION 0.77; micro-average 0.92. Name and place detection is best-effort; deny-lists help, and an optional LLM tier is coming to hosted plans.
Known limits
- The hosted API sees plaintext in memory during each request.
- Key-only restore transforms anything that looks like an
fpetoken: if an LLM invents a new card or email in its reply, restore will turn it into a different value. Names and places are not affected (authenticated tokens). - Name and place recall is about 0.80 on the synthetic benchmark.
Compliance status
No SOC 2 or ISO 27001 certification today. Compliance work is in progress. We are glad to walk your security team through the design and answer questionnaires — contact us.
Planned — not available today
- Self-hosted / in-VPC edition (design-partner pilots)
- Hosted LLM detection tier