AI dashboard privacy: what can it see, send and change?

AI dashboard privacy comes down to three questions: what can the app access, what does it send to another service, and what can an assistant change? A local note, a synced note and a note included in a cloud AI prompt can follow different paths. Treat each path as a separate decision.

You do not have to become a security engineer to make those decisions. You do need more information than a lock icon and the word “private.”

Separate the layers of access

This table is a way to evaluate products, not a claim that every app implements the same boundaries. Read the documentation for the specific release you use.

A visible page is not the same as shared context

An assistant beside a webpage might receive selected text, a captured page, attached files or none of those until you ask. Its position on the screen does not establish its access.

Look for a context preview or a clear description of what is included before Send. If you only need help rewriting a paragraph, the whole page may be unnecessary. If the page contains names, account details or private messages, a small excerpt is easier to review.

The same principle applies to notes. “Use this note” should have a recognizable target. “Use everything you know about me” is harder to audit and may include more information than the task requires.

Connected accounts need their own review

Signing into the dashboard is not the same as authorizing it to read a calendar. A Google sign-in and a request for Calendar data can appear in related flows but serve different purposes.

Google's third-party account access guide explains reviewing and removing access. Use the provider's permission screens and check which account you are connecting, particularly if you have both personal and work accounts.

Revocation prevents future authorized access through that grant; it should not be treated as proof that all copies of previously shared information have been deleted. Check the receiving product's deletion policy separately.

For a calendar-specific example, see Google Calendar on a new tab dashboard.

MCP connects tools; it does not settle every privacy question

The MCP architecture describes clients connecting to servers that expose capabilities such as tools and resources. That common interface is useful, but it does not make every server equally trustworthy.

Before connecting a dashboard to an assistant, establish what the connection can read and what it can write. Ask whether the available tools are limited by the account and permissions you granted. Review how the product handles consequential actions and failures.

A read tool can still expose sensitive information. A write tool can alter saved work. Choosing read-only access reduces one kind of risk without making every possible read appropriate.

Our ChatGPT connection guide gives a small read-first trial for checking a supported integration without starting with broad changes.

Keep credentials out of conversations and widgets

Do not paste session tokens, passwords or private API keys into a chat because an assistant says it needs them. Use the application's documented authentication or secret-management flow.

The same applies to a custom widget. Code delivered to your browser can be inspected, even if it is compressed or difficult to read. Minification is not a way to protect a secret embedded in that code.

A generated widget also needs a clear description of its data use. A counter may need only a small amount of stored state. A widget showing live private records needs an authorized integration. A prompt cannot create permission where none exists.

Start with a low-stakes widget and inspect its behavior. The custom-widget guide includes a practical preview sequence.

Ask what “local AI” means for the feature you use

An on-device model can process supported work locally. That does not establish that the whole app is offline: account sync, external widgets, analytics and cloud features can have separate network activity.

Likewise, an app offering cloud AI does not mean every click sends content to a model. Ask about the particular operation. Generating a widget, clicking its counter and syncing its stored value may be three different processes.

Avoid buying on an absolute claim when a feature-level explanation would be more useful. A good product should make those boundaries understandable.

Dashy's direction and its current boundary

Dashy V2 is developing local dashboard essentials alongside optional AI, connected services and MCP tools. Its AI guide describes preview controls, selected context and account requirements. V2 is planned for October 2026 and is not generally released; these statements should not be applied automatically to the original store version.

We also distinguish stopping work from undoing it. Canceling an AI request cannot reverse a change that already completed. Supported undo and deletion controls have their own limits, which should be checked for the action involved.

A small review before your first real task

Connect only the service you need. Read the permission request. Try a non-sensitive note or disposable task. Confirm the result in the destination app, then find the disconnect control before you need it.

For work material, follow your organization's rules even if the integration is technically available. For personal material, share the smallest useful context. You can always add a missing detail; you cannot make information unshared simply by removing it from the next prompt.