Locke, by Sonomos · Coming Soon

SECURITY

Security built into how Locke works, not bolted on after.

Detection runs on your device. Masking is one-way. Network interception is scoped to exactly what Locke protects, and nothing else. This page covers how, in general terms — for the deeper technical documentation, see below.

For company-level trust info — independent scans, our disclosure policy — see Security & Trust.

Detection Runs on Your Device

Every scan — structural, contextual, and free-form — happens locally. Nothing is sent out to be analyzed.

Masking Is One-Way

Sensitive values are replaced with a non-reversible placeholder. There's no key that could ever reconstruct the original.

Scoped, Not Total, Interception

Locke is designed to scope inspection to the specific AI tools it protects, rather than act as a general-purpose VPN or traffic monitor.

Build inputs you can trace

Release inputs are pinned to specific component versions, making changes reviewable instead of dependent on whatever happens to be latest.

DETECTION & MASKING

What happens to your data

Every step below happens on your device, before anything is sent anywhere.

Three ways to find sensitive data

Structural validation for data with a checkable format (like a checksum), contextual matching for data identified by what's around it, and free-form recognition for names, organizations, and places with no fixed pattern.

A placeholder, not a fake value

Detected values are replaced with a short, consistent token before anything leaves your device — not a realistic-looking substitute. The same value stays consistent within the document being processed, so that document remains coherent.

Memory-safe by construction

Selected Rust modules use explicit safe-code boundaries, reducing the risk of memory-safety mistakes in those modules while the wider application still uses defense-in-depth.

Nothing sensitive in the logs

What Locke records about a detection is metadata such as its type, length, and confidence rather than the detected value itself.

Reversible tokenization — restoring a reply so it shows your real values automatically — is planned for a release shortly after Locke 1.0. It is not part of Locke 1.0. Today, masking is one-way: there's no way to reconstruct your original value from what Locke sends.

NETWORK

How Locke talks to the network

Locke inspects traffic to the AI tools it protects so it can screen it locally before it goes out. That inspection is deliberately narrow.

A narrow, specific scope

Locke only intercepts traffic to the AI-tool domains it's built to protect. General browsing, email, and everything else passes through as an untouched, still-encrypted connection.

Locally generated trust material

Locke creates local certificate material for its active gateway session. Nothing is shared across users or devices, and the private key stays on your machine.

The private key stays local

The local gateway's private key is not transmitted. Trusting its local root is an explicit device-level decision that enables inspection only for the supported AI flows Locke is configured to protect.

A fully verified connection onward

After inspection, your request continues to the real AI provider over a standard, fully verified TLS connection — the same trust checks your browser would normally perform.

Blocks what it can't verify

For supported prompt bodies, if Locke cannot confirm that screening completed, the gateway can fail closed rather than guess and pass that content through.

THE APPLICATION

How the app itself is hardened

The privacy engine is only half the story. The app around it is built to reduce its own attack surface too.

A locked-down interface layer

The app's interface runs with context isolation on and no direct access to system APIs. It can only reach the rest of the app through a small, explicit, allow-listed bridge — not a wide-open channel.

No third-party analytics or telemetry

Locke doesn't bundle any third-party analytics or crash-reporting SDK. Any first-party usage data is opt-in, limited to coarse counts, and never includes your content.

Everything stays local

Preferences and detection history are stored on your device. No account is required to use Locke, and nothing about your usage is synced to the cloud.

BUILDS & RELEASES

Built and shipped deliberately

Security isn't only what the app does at runtime — it's how confidently we can say what shipped is what was reviewed.

Pinned build inputs

Release artifacts are assembled from specific versions of their components, so the inputs can be inspected and reproduced.

Deliberate version changes

Updating a component is treated as a deliberate release decision, with the detailed review process available through the Trust Center.

Nothing ships without checks passing

Formatting, linting, tests, and documentation integrity are treated as release checks; detailed release evidence is available through the Trust Center.

Release signing is part of the production plan

Production Windows release signing is planned so users can verify provenance and integrity. Current development builds should not be treated as signed releases.

Dependency risk is reviewed

Locke's dependencies are reviewed as part of release work, with deeper vulnerability-management details maintained in the Trust Center.

HOW THIS COMPARES

Why on-device changes the security story

Most of what makes Locke secure follows from one decision: processing never leaves your device. See the full comparisons for how this plays out category by category.

Nothing server-side to breach

Cloud-based privacy and DLP tools inspect your original content on someone else's server. Locke does not send your original sensitive values anywhere to be analyzed, so there is no Sonomos-side copy of those values to breach, subpoena, or leak.

No key that unlocks your data

Tools built on reversible encryption keep a key somewhere that can reconstruct your original values. Locke's masking in 1.0 is one-way — there's no key to steal, because there's no way back to the original from what's sent.

A narrow blast radius by design

General-purpose proxies and VPNs typically see everything you do online. Locke is scoped to the specific AI tools it protects, so what it can ever see is deliberately limited, not broad by default.

Go deeper

This page covers the general picture. Two places to go for more, depending on what you need.

Sonomos Support Center

User guides, how detection works in plain terms, and light security discussion — open to everyone.

Visit the Support Center

Sonomos Trust Center

Detailed architecture, security practices, and compliance documentation for customers and prospective customers — access granted on request.

Open the Trust Center

Locke is in development. Request early access and we'll be in touch as soon as builds are available.

Request early access