SP Vault — threat model
Scope: the personal zero-knowledge vault (web, macOS, iOS, API on MacLustr). The managed-secrets subsystem
(server-side secret use for agents and brokered operations) has a different, weaker model, see
services/managed-secrets/THREAT_MODEL.md. Protocol reference: docs/crypto-spec/SPV1.md.
Assets
- Vault plaintext (credentials, TOTP seeds, notes, cards, documents, attachments, names, tags).
- Key material: Account Root Key (ARK), account keyset, vault keys, recovery key, device private keys.
- Integrity and availability of the vault (no silent loss, rollback or substitution).
- Metadata (who, when, how much).
Trust assumptions
- The user's unlocked device and its OS are trusted while in use. Nothing in this design survives a fully compromised unlocked endpoint.
- Apple Secure Enclave, CryptoKit, WebCrypto and the pinned dependencies behave as specified.
- The server (MacLustr node, tunnel gateway, database, backups, operators) is not trusted for confidentiality or integrity of vault content; it is trusted only for availability.
Adversaries and outcomes
| # | Scenario | Defences | Residual risk |
|---|---|---|---|
| T1 | Malicious or compromised server / database or backup theft | Client-side encryption; server holds only ciphertext, public keys and SHA-256 of the recovery authenticator; device list is certified by the account signing key so the server cannot enrol a device; AAD binds every record to vault/id/type/revision. |
Metadata in SPV1 §8. Denial of service, deletion, withholding of recent writes, rollback of a never-synced device. Traffic analysis. |
| T2 | Malicious server serving hostile JavaScript to the web app | Strict CSP (no inline script, no third-party origins), Trusted Types, SRI-style build manifest with published hashes, reproducible build, deployment only through mld from a signed git commit. |
Not eliminated: a server that controls the HTML controls the web client. The native apps are the high-assurance clients; the web app says so in Settings → Security. |
| T3 | Stolen locked device | ARK is never stored unwrapped; unwrapping needs the Secure Enclave key, which requires biometrics or the device passcode; the local database holds record ciphertext only (never decrypted content) plus sync metadata in clear (vault/record ids, member handles and public keys, cursors); data-protection class …WhenUnlockedThisDeviceOnly; remote revocation removes the envelope and sessions. |
An attacker who knows the device passcode can unlock (policy userPresence; strict biometric-only mode is an option). Revocation does not reach an offline device. |
| T4 | Stolen unlocked / compromised endpoint (malware, forensic image of a running app) | Auto-lock on inactivity, on sleep and on background; secrets zeroed on lock where the runtime allows; no plaintext on disk outside the encrypted store; hardened runtime, sandbox. | Full compromise of an unlocked endpoint discloses everything it can decrypt. After such an event: revoke the device and rekey the account (SPV1 §6), then rotate the stored credentials themselves. |
| T5 | Phishing of the server login | Passkeys only (origin-bound, no password to type); no SMS/e-mail fallback. | Social engineering to approve a device pairing: mitigated by the tag check + 6-digit comparison + explicit device details; cannot be fully prevented. |
| T6 | Phishing of stored credentials (lookalike domains) | Fill only on exact registrable-domain match computed with the Public Suffix List, never on http: for a saved https: login; IDN shown in punycode when scripts are mixed; lookalike warning (confusable skeleton vs saved hosts); destination origin shown before filling. On Apple platforms matching is performed by the OS for the AutoFill extension (service identifiers). |
A user can still copy a password manually into a hostile page. |
| T7 | Browser extension compromise / hostile extension in the web vault's page | No extension is shipped in v1. CSP + Trusted Types reduce injection; secrets are rendered only on explicit reveal. | An extension with full page access can read whatever the page shows or holds in memory. Use native clients. |
| T8 | Supply-chain attack | Minimal dependencies, exact pins + lockfile with integrity hashes, pnpm audit, SBOM (CycloneDX), secret scanning, no post-install scripts (ignore-scripts), Swift uses only Apple frameworks. |
A malicious release of a pinned dependency adopted during an update. Updates are reviewed diffs. |
| T9 | Brute force | No password anywhere: keys are 256-bit random. Recovery authenticator is 256-bit; endpoints are rate-limited per IP and per account regardless; pairing codes are 128-bit, single use, 5 min. | — |
| T10 | Replay / request forgery | WebAuthn challenges single-use with expiry and origin + RP ID verification; session tokens random 256-bit, stored hashed, rotated on privilege change; SameSite=Strict __Host- cookie + custom header + Origin check for the web; idempotency keys; base_rev on writes. |
— |
| T11 | Rollback / stale state | Client-side high-water marks for seq and per-record rev; AAD binds rev; state digest comparison between devices. |
SPV1 §7 limits (freshness for an offline or new device). |
| T12 | Sync conflict causing data loss | Server never overwrites on stale base_rev; the client keeps the losing version as a conflict copy that the user reviews; previous revisions retained in history; trash with 30-day retention. |
— |
| T13 | Recovery abuse | Recovery needs the 256-bit recovery key (never on the server). Recovery events are audited, shown on every device and end all other sessions' pending states; RK rotation is offered after each use. No operator backdoor exists. | Anyone holding the printed recovery kit owns the vault. Losing every device and the kit is permanent loss, stated during onboarding. |
| T14 | Cross-tenant access (IDOR, enumeration) | Every query is scoped by the session's account and by server-side membership rows; identifiers are random UUIDs; uniform errors and timing for unknown handles; tests in tests/security. |
— |
| T15 | Malicious agent / prompt injection (SP Vault for Agents) | Agents never get the personal vault. They hold a scoped identity; each grant names one secret, an action class (operate vs read), a lease TTL and a use budget; high-risk actions wait for a human approval made in an authenticated client; all tool output is treated as data; kill switch revokes every lease at once; hash-chained audit log. |
A secret delivered in read mode can be disclosed by its recipient. Brokered operate mode is preferred and labelled. |
| T16 | Insider / operator on the cluster | Same as T1 for the personal vault. For managed secrets the operator of the node can decrypt (key on the node, mode 0600, separate from backups): stated in the UI wherever a secret is moved there. | By design for managed secrets. |
| T17 | Leak through logs, crash reports, notifications, clipboard, screenshots | Server logs carry request id, route template, status, duration only; no bodies. Clients: no analytics, no third-party crash reporter; app-switcher privacy cover; clipboard marked concealed/local-only with expiry where the OS supports it and cleared after a delay best-effort. | Clipboard managers and universal clipboard may already have copied the value. |
| T18 | Malicious ciphertext / parser attacks | Strict schema validation before and after decryption; size limits; fuzz tests on envelope and import parsers; AEAD failure is fatal for that record only and surfaced. | — |
Out of scope
Rubber-hose coercion, hardware attacks on the Secure Enclave, a compromised Apple ID that syncs the login passkey (the passkey alone gives no vault access — a certified device or the recovery key is still required), nation-state traffic correlation.
Launch gates
Independent security review of SPV1 and both implementations, external penetration test of the API and web app,
real-device tests of Keychain / biometrics / AutoFill, and a restore drill are required before real credentials
are migrated. Their status is tracked in SECURITY.md.