Security & Trust
What we operate today, what we have not done yet, and how you can check our work without taking our word for it.
Compliance status
SOC 2 Type I readiness in progress; not attested. No auditor has been engaged and no SOC 2 report exists. We are documenting controls and collecting evidence so that an audit could start.
CertifiedData does not claim SOC 2 (any type), FedRAMP, HIPAA, CMMC or ISO 27001 status, and is not registered in SAM. A CertifiedData certificate is cryptographic evidence of a dataset's origin and integrity; it is not a compliance determination. See compliance documents.
Controls we operate
Certificate signing
- Certificates are signed with Ed25519; dataset fingerprints use SHA-256 with RFC 8785 canonicalization.
- The production signing key is supplied to the API as a deployment secret (environment variable), not stored in the database. The certificate_signing_keys table holds public keys only.
- The HMAC-SHA256 fallback signer is refused when the API runs in production.
- Keys can be marked revoked (revoked_at / revocation_reason); revoked keys fail verification and appear in the published registry.
- Limits: the key is not held in an HSM or cloud KMS, and a documented rotation schedule does not yet exist.
Access control
- User authentication through Supabase JWTs (JWKS-verified); server-side sessions use httpOnly, secure cookies.
- API keys are random (192 bits) and only their SHA-256 hash is stored.
- Admin routes require an email allowlist and a database admin flag.
- Rate limiting on authentication, API and public verification routes.
Application and transport
- HSTS, X-Frame-Options DENY, X-Content-Type-Options nosniff, Referrer-Policy and a Content-Security-Policy on API and web responses; Cloudflare sits in front of the site.
- Zod validation on request payloads and request size limits.
- Known limitation: the CSP still allows 'unsafe-inline' scripts (tracked as finding F-02).
Tamper-evident records
- Database triggers make the certificate transparency log, audit vault records and audit logs append-only (UPDATE/DELETE are rejected unless an explicit maintenance override is set in a transaction).
- Certificate transparency log entries are hash-chained; each new append verifies the previous entry and fails closed if it does not match.
- A nightly job recomputes every entry hash and link in the log.
Engineering process
- CI runs CodeQL, dependency review, a security audit workflow (dependency audit, secret scan, workflow lint, route inventory, header baseline, security regression tests) and a trust harness over certificate and verification code.
- Dependabot proposes dependency updates weekly.
- Security findings are tracked in a written findings register, including unresolved ones.
- Limits: no independent penetration test has been performed; formal policies are drafts that have not yet been approved.
Verify a certificate yourself
- Compute the SHA-256 hash of the dataset.
- Compare it with the dataset hash in the certificate.
- Verify the Ed25519 signature with the public key from /.well-known/signing-keys.json (includes the revoked-key list).
The public verifier at /verify needs no account. Specifications: verification spec, signing keys, certificate policy.
Transparency log status
| Live log tail | index 4996, appended 2026-11-20T23:15:39.264Z |
| Last full-chain audit (dated) | November 13, 2026: 4,464 entries, 0 hash mismatches, 0 linkage breaks |
| Audit cadence | Nightly, read-only; this dated figure is updated by hand |
The live tail is read from the public /api/log/latest endpoint and shows only that appends are happening; it does not itself prove the chain is intact. The audit line is the result of recomputing every entry hash and link on the date shown. Entries can be browsed at /transparency/certificates.
Incident history
April 2026 — chain break at entry 546. On April 8, 2026 the certificate transparency log stopped verifying at entry 546. Our append-time check fails closed, so certificate issuance halted until the chain was repaired. It was noticed on April 23, about 15 days later, by a user reading the public log, not by our monitoring.
What changed since: Telegram alerting on the append-time check, a nightly full-chain audit, database append-only enforcement and serialized appends. The nightly audit was found in November 2026 to exit successfully even when it found bad rows; that is fixed in our repository and will take effect on the next deployment.
What we do not know: the root cause of the original corruption is not proven, and the repair was not recorded in enough detail to publish. The chain has verified clean on every nightly run whose logs we still hold (August 16 – November 13, 2026). Details: transparency log.
Responsible disclosure
Report vulnerabilities to [email protected]. Include a description, reproduction steps, impact, and optionally your name for credit. We aim to acknowledge within 24 hours and provide an initial assessment within 72 hours. Please do not access other users' data, run denial-of-service tests against production, or disclose publicly before we have had reasonable time to fix.
Last updated: November 13, 2026