Secure input
The supported AI page does not receive the raw draft. It receives only the helper-protected version selected for sending.
Trust and documentation
Know where private information goes.
Connect protects content locally, but the protected content you choose to send still goes to your selected supported AI service. The product reduces exposure; it does not make an AI request local. ChatGPT and Claude are currently supported in Chrome and Edge, with more AI providers being added.
Details when you need them
Start with the plain-language trust notes.
The sections below explain the important boundaries. Use these references when you need implementation details, policies, or support.
The boundary
Raw text stays in the secure composer.
With Protection Mode on, the original draft and original file bytes remain inside an extension-origin frame until the local helper returns protected content.
Local engine
Detection, templates, mappings, file protection, and restoration run in the native Windows or Mac helper rather than the inspectable extension shell.
Encrypted continuity
Each conversation has an AES-256-GCM encrypted vault. The installed helper protects the master key with macOS Keychain or Windows per-user DPAPI.
Isolated restoration
Restored transcript views render in extension-origin frames. Restored values are not written back into the supported AI page.
What crosses the network
The protected prompt still leaves your computer.
Sent to your AI service
The protected prompt, protected file copies, and normal account interactions are sent through the supported AI service you use and are governed by its terms, settings, and retention choices.
Not sent by Connect
The restore vault, original-to-alias mappings, secure credential key, raw local logs, custom template definitions, and original files are not part of the AI request.
Limited subscription check
The helper contacts the CloakShift service about once per day to verify a signed subscription certificate. That check includes account and registered-computer metadata, but never prompts, files, mappings, restored content, or activity logs. There is no CloakShift product-analytics or remote prompt-logging endpoint.
Website hosting
Like most hosted sites, cloakshift.ai and its infrastructure may process standard request information such as IP address, user agent, route, and timestamp for delivery and security.
Evidence
Tests reduce risk; they are not certification.
Measured with Maximum protection in the frozen V10 Connect benchmark.
Strict exact-value recall across 44 synthetic prompt and file cases.
No complete annotated value was silently exposed.
Every expected protected value restored exactly.
Known risks
What Connect cannot eliminate.
Detection failure
Automatic detection can miss names, identifiers, context, metadata, and domain-specific values. Review protected content before sending it.
Inference from context
Distinctive facts can identify a person or organization even after direct identifiers are replaced.
Model transformation
An AI can rewrite, merge, remove, or hallucinate aliases. Restoration cannot recover a reliable match that no longer exists.
Local compromise
Malware, another local user, screenshots, clipboard history, or an unlocked account can expose originals and vaults.
Page compatibility
Supported AI sites can change their page structure. Connect fails closed where possible, but every integration requires ongoing compatibility testing. ChatGPT and Claude are currently supported in Chrome and Edge, with more AI providers being added.
Compliance responsibility
Connect is not a legal opinion, security audit, compliance certification, or substitute for authorization, policy, contracts, and professional review.