SECURITY & ENTERPRISE · UPDATED JULY 2026
Security review, in plain facts.
Everything an IT or InfoSec team needs to evaluate LockedIn Flow for a corporate deployment: where data goes (nowhere), what it touches (almost nothing), how it's signed, and how to roll it out.
The one-sentence posture
LockedIn Flow has no vendor backend — no server of ours ever receives audio, transcripts, vocabulary, screen content, or telemetry. The largest breach surface in this category simply does not exist here.
Data flow (complete)
| Stage | Where it happens | Leaves the device? |
|---|---|---|
| Microphone capture | In-process, 16 kHz mono, in-memory | No |
| Speech recognition | Parakeet TDT v3 on the Apple Neural Engine | No |
| Text cleanup | Rules engine + Apple FoundationModels, on-device | No |
| Insertion | OS APIs (Accessibility / pasteboard) | No |
| History & vocabulary | AES-256-GCM files, Keychain-held key | No |
| Model download | One time, ~600 MB, huggingface.co | Outbound, once |
| Update check | Optional, HTTPS, Ed25519-signed appcast | Outbound, optional |
Permissions — the whole list
- Microphone — capture only.
- Accessibility — to insert text at the cursor and to refuse password fields. Screen content is never read, stored, or transmitted.
- Never requested: Full Disk Access, Contacts, Calendar, Location, Screen Recording, Input Monitoring, Camera.
Code signing & supply chain
- Signed with Developer ID Application: Zero G Foundry LLC (643FW3ZH6M), Apple hardened runtime, minimal entitlements (audio input only).
- Notarized and stapled via Apple notarytool — Gatekeeper-clean install.
- Updates: Sparkle 2, Ed25519-signed appcast items, HTTPS only.
- License validation: offline Ed25519 signatures. No activation server, no phone-home, no device fingerprinting.
- Dependencies are all MIT/Apache-2.0 (Parakeet model: CC-BY-4.0). No GPL. SBOM available on request.
Offline enforcement for locked-down fleets
After the one-time model download, the app works with all egress blocked —
verified in our self-test with network fetches refused at the library level. Models can be
pre-seeded via MDM into ~/.cache/fluidaudio, and there is no account, activation,
or revalidation heartbeat for a firewall to learn to hate.
Data governance
- Raw audio is never written to disk, ever.
- Transcript history: AES-256-GCM at rest, retention from off to forever, user-visible purge. An enterprise policy file can pin retention to zero.
- Logs are metadata-only; transcript content is never logged.
- No analytics, no crash-report endpoints, no third-party SDKs with network behavior.
Deployment
- Individuals: notarized DMG, drag-install, guided first-launch permission flow.
- MDM fleets: deliver the signed DMG; PPPC payloads can pre-approve Microphone + Accessibility; models pre-seedable; firewall can be closed post-provisioning.
- VDI: works subject to microphone redirection in your VDI stack — the same caveat as any dictation tool.
- Windows: in development with the same posture (Rust/Tauri core, Parakeet via sherpa-onnx, same offline licenses). Join the waitlist.
What we don't claim
No HIPAA, FedRAMP, or SOC 2 certification is claimed. The architecture removes the largest compliance surface — no third party ever receives data — and your own device management (FileVault, screen lock, MDM) completes the posture. We're happy to answer your security questionnaire: security@lockedinflow.com.
FREE 14-DAY TRIAL · NO CARD · NO ACCOUNT
Pilot it with your team this week.
Everything unlocked for 14 days. Nothing to configure, nothing to whitelist except one model download.
Download for macOSArchitecture statements verified against the shipping build, July 26, 2026. The full whitepaper (threat model, dependency inventory) ships inside the app repo and is available on request.