// TRUST AND DATA BOUNDARIES
Understand what leaves your device.
DMZ’s detection and restoration mapping run on your device. It sends neither the prompt nor the mapping to DMZ servers. The AI site receives whatever text is actually submitted, including missed identifiers and anything sent unchanged.
LOCAL PROCESSING · CHECK FOR YOURSELF
Evidence behind local processing.
This release processes text on your device using the bundled model. DMZ uses no remote inference or prompt/mapping uploads to DMZ servers. Here you can inspect what was tested and repeat the check.
0.1.3 alpha.3 · Tested 2026-10-27 · Chromium 156.0.8078.4
What did the test show?
We tested the same Alpha 3 package offered for download. In a fresh Chromium profile, with the browser offline and synthetic ChatGPT/Claude pages, local name detection, masking, restoration and pause/settings worked. Only two document requests were observed; both were fulfilled locally by the test. No unexpected HTTP(S) requests or page WebSocket attempts were observed in these workflows. A separate control confirmed that the recorder detected a deliberate network attempt.
Inspect the permissions
The ZIP’s manifest.json declares offscreen (local scanner) and sidePanel (settings). Content scripts match chatgpt.com and claude.ai only. No host_permissions or externally_connectable are declared. Extension pages use connect-src 'self', allowing their own packaged resources. This is not a blanket network ban on content scripts, and none of these settings alone proves data leakage impossible.
Extension network requests · Chrome documentation ↗
Content-script security policy · Chrome documentation ↗
Repeat the check
- Download the Alpha 3 ZIP and compare its SHA-256 with the value below. Extract it into a fresh folder without changes.
- Download the verification script. In an empty folder with Node 22 or later, install Playwright and Chromium using the commands below. Tool installation needs internet; the workflow test then runs offline.
- Run the script with paths to the ZIP and its extracted extension subfolder. It checks release identity, uses a temporary browser profile and writes report.json. Inspect the script yourself; it uses fabricated text only.
npm install playwright@1.63.0
npx playwright install chromium
node verify-local-processing.mjs /path/to/dmz-chromium-extension-0.1.3-alpha.3.zip /path/to/extension report.jsonSHA-256 identifies the file; it does not establish its safety:6e59e15141a251bee8fb0300acb2ff681bba0d213d744ee264e7a94a4ba66c91
What this evidence covers
This is a bounded maintainer test, not an independent audit or proof that networking is technically impossible. It does not include OS packet capture, a DNS/QUIC audit or all browser background traffic. It uses mocked pages, not authenticated AI services. It applies only to this release and must be repeated for a new package.
ChatGPT and Claude can see the original composer draft and handle their own network traffic. Paused masking, detection misses and unchanged attachments can expose information through the AI site. Links you open from settings visit DMZ/Kastro websites. dmz.is itself requires network access; this is not a claim that your browser or AI website is offline.
You can also inspect Network in DevTools for the service worker and extension pages from chrome://extensions, plus the content script in the chat tab. Start recording before testing and use fabricated data only. One empty Network panel does not cover all browser processes. An independent audit would add assurance; none has been completed.
Before submission
The original draft is typed into the AI site’s composer before DMZ masks it. Page scripts can already read it or observe input events. DMZ does not isolate the composer or protect against pre-submit telemetry or autosave.
A local aid
Supported text passes through local detection and masking. A per-tab map links selected original values to visible markers. Unchanged returned markers can be restored in a local extension view; the AI page retains its masked reply. Resetting, reloading or closing the chat tab clears that session.
Best-effort submission
DMZ does not enforce a submission block. Scanner, masking or composer failures produce warnings and submission continues. A trained-scanner failure falls back to basic masking, reducing name coverage. Files, images and attachments are sent unchanged with a warning. Check the text yourself before sending.
Detection has limits
Names, addresses and other identifiers may be missed or innocent text masked. Context can still identify someone. Site changes or unsupported send paths can bypass interception. This preview is for fabricated-data testing and is not a complete security boundary.
Your AI service
ChatGPT and Claude are services you choose independently. Their account, retention and data-use terms apply to submitted content. DMZ does not promise provider zero retention, anonymity, legal compliance or certification.
Evidence
The benchmark page shows reviewed development summaries and limitations. Independent validation is incomplete. A free release must still meet the existing quality and no-regression requirements; becoming a community project does not prove safety.
EXPLORE
Updated 27 October 2026.