Trust Center
Security whitepaper: the pilot, its controls and its open findings
Version 0.5.0, October 2, 2026. Written from the developer’s own review. It is not an audit, a certification or an independent assessment.
1. Scope and honesty
This paper covers the synthetic-data pilot hub and its clients. It does not cover a production deployment, because none exists. The reviews behind it (one for each of Waves 1 to 4) were performed by the developer, not by an outside party. There has been no penetration test, no fuzzing campaign, no dependency vulnerability scan and no review of the cloud account configuration. Its own conclusion is that the pilot is safe to demonstrate with synthetic data and must not hold real patient data until the findings in section 4 are addressed and an independent review is done. Every external system in the pilot (laboratory, pharmacy, payer, FHIR and HL7 partners) is a simulator, so no real external attack surface has been tested.
2. Design model
Access is deny by default and derives from the role roster (1,650 roles in tiers T0 to T9 and TS), care-relationship scope and, in the design, purpose and consent. Privileged actions need a second person. Audit is designed to be tamper-evident. What each role may do is a server-enforced permission matrix (23 role classes and 103 actions). The on-site hub and its signed, encrypted sync are built; the larger design (a ledger of hashes, key hierarchy, sealed terminals) is described in the architecture deep-dive and is not built.
3. Controls in the pilot
| Area | Control | Status |
|---|---|---|
| Authentication | Passwords of at least 12 characters (14 in the Hospital preset), forced change at first sign-in, one-time temporary passwords, TOTP MFA with recovery codes (on by default for the privileged tiers T7 to T9), replay protection, lockout after 5 failures with a security alert, generic errors, idle and manual lock, session list and revoke | In the pilot (v0.5.0, synthetic data) |
| Authorisation | Deny-by-default tiers, care-team scope, tenant isolation tests, two-person approval for privileged user changes, sysadmin roles without chart access | In the pilot (v0.5.0, synthetic data) |
| Cryptography | HTTPS and secure WebSockets; field-level AES-GCM for national IDs, MFA secrets and registration history; sessions stored as hashes | In the pilot (v0.5.0, synthetic data) |
| Application security | Security headers and CSP on API and static pages, CORS allow-list, parameterised SQL and no raw-markup APIs (both unit-tested), Electron sandbox with three IPC channels | In the pilot (v0.5.0, synthetic data) |
| Configuration integrity | Signed manifests, locked safety floors re-applied at run time, tamper detection on packages, dual-operator publishing | In the pilot (v0.5.0, synthetic data) |
| Audit and emergency access | Hash-chained audit with verify; break-glass with reason, time box, alert and mandatory review | In the pilot (v0.5.0, synthetic data) |
| Permission matrix | 23 role classes by 103 actions, enforced on the server, editable only through the signed manifest within a floor and a ceiling; 8,565 allow and deny checks on synthetic roles | In the pilot (v0.5.0, synthetic data) |
| Separation of duties | Capturing a charge, managing a claim, submitting it and posting payment are separate actions; only the named prescriber can sign a prescription; controlled substances need a password step-up | In the pilot (v0.5.0, synthetic data) |
| Patient portal isolation | A separate patient identity and sessions; activation codes are one-time, last 72 hours and die after 5 tries; staff and portal tokens are refused on each other’s routes | In the pilot (v0.5.0, synthetic data) |
| Sync and backup integrity | Replication bundles are AES-256-GCM encrypted, signed per site, hash-chained and sequence-numbered; backups carry per-table digests and a signed manifest and are proven by an isolated restore drill | In the pilot (v0.5.0, synthetic data) |
| Deploy gate | After each deploy a live self-test of every piece runs on hidden probe records; a failure blocks the release and raises a system alert | In the pilot (v0.5.0, synthetic data) |
| Decision support and printing | Every decision-support override needs a reason scaled to severity and is audited and reviewed; prints are logged, watermarked, rate-limited and need a step-up for protected compartments | In the pilot (v0.5.0, synthetic data) |
4. Open findings
Severity labels are the review’s own.
| Ref | Severity | Finding | What closing it needs | Status |
|---|---|---|---|---|
| F2 | Medium | The demo vendor signing private key is in the demo code base, and the demo hub trusts its public half. | A production key custody model (hardware-backed, split custody) and rotation. Custody is undecided. | Open decision |
| F3 | Medium | Passwords are hashed with PBKDF2-SHA256 at 100,000 iterations rather than Argon2id, and there is no breached-password check. | A move to Argon2id (the specification asks for 64 MiB memory and t≥3 or an equivalent) and a breached-password screen. | Open decision |
| F4 | Medium | The field-encryption key is a real hub secret, but there is no rotation, no per-tenant key and no key management service. A local default is insecure. | The key hierarchy in the design, with per-tenant keys and scheduled rotation. | Open decision |
| F8 | Low | Privacy compartments are flags plus an attested purpose, not a legal consent engine. | A consent module and privacy operations, with local legal review. | Open decision |
| F11 | Info | The audit chain is tamper-evident but not write-once, and has no external anchor. The database owner could rewrite the chain and its head. | Append-only storage and anchoring, as in the design. | Open decision |
| F12 | Info | There is no per-IP rate limit; only per-user limits and account lockout. | Edge rate limiting and abuse controls. | Open decision |
| D-028 | Medium | The HMAC secrets that sign on-site sync, the backup-manifest signing key and the vendor key are demo-grade: they are kept in the code base or the database. | An HSM or key management service and a custody decision before any real use. | Open decision |
| S-live | Info | One of the 27 live click-through steps of Wave 4 failed: a too-short decision-support override reason showed no message in the dialog in the headless test browser on the live site. It passes locally, and the server refuses a short reason (a 409 response), but the on-screen message on the live site has not been shown to work. The cause is unresolved: a test timing issue is suspected, a product defect is not ruled out. | Find the cause and show the on-screen message working on the live site. | Open decision |
Findings from the Wave 2, 3 and 4 reviews that were fixed include an HL7 escaping bug, a replication case where independently created duplicate records could silently diverge (now a conflict that pauses sync), and race conditions in four self-tests. Other items were accepted for the pilot: the break-glass lookup reveals that a restricted record exists (audited, role-limited); CORS allows any localhost origin; bearer tokens live in session storage; a hosting analytics beacon is blocked by the CSP; login timing was not tested statistically. Static-page headers (F1) were missing and are fixed.
5. Verification evidence
Run by the developer between September 30 and October 1, 2026, on synthetic data, and listed here without being combined into one figure. Unit tests: 137 of 137 pass (re-run on October 2, 2026). End-to-end suites on the live hub: 108 (demo), 201 (Wave 1), 64 (packaging), 114 (Wave 2), 93 (Wave 3) and 85 (Wave 4) checks, none failed. A permission-matrix run of 8,565 checks with none failed. Security probes: 34, 105, 96 and 58 checks for Waves 1 to 4, with no breach. Click-throughs of 37, 25, 28 and 26 of 27 steps (one step is the open item above). A scan of 427 screens with no problems. On-site hub end-to-end: 37 of 37. The server circuit was green with all 36 pieces on three tenants and 25 on the Small Clinic tenant. These are internal tests. They are not an audit and do not show suitability for real patient data.
6. Designed controls
Closed on-site servers beyond the current on-site hub; a federated ledger that holds only hashes, consent records and audit events; a key hierarchy with HSM custody; a read-side limiter, safe-degrade sync and a scroll-rate lock; sealed terminals; three-stream server-to-server transfer with hybrid post-quantum key exchange; keystroke-cadence analytics as a review flag only; crypto-shredding for deletion. All are Designed. Several parameters are open decisions, and hardware effectiveness is untested.
7. Not done
- No independent penetration test or code review; scope and provider are undecided.
- No certification, attestation or compliance claim of any kind.
- No production key management, no off-site or immutable backup, no tested disaster-recovery plan beyond the isolated restore drill, no monitoring programme.
- No privacy-operations or breach workflow, no application-security program and no compliance documents.
- No external pentest, no fuzzing, no load or chaos testing and no dependency scan.
- No operating-system or cloud-account hardening review.
8. Reporting a vulnerability
Email [email protected]. See security.txt. There is no bounty programme. The pilot holds only synthetic data.
Send us your security questionnaire
We will answer plainly, including “not yet”.