DictaFlow

November 15, 2026

Lovable Data Exposure: Check Saved Work Notes in 2026

Check access to the summary, not only the original message.

A locked notes cabinet with an open side allowing unmarked paper copies to escape
Original procedural editorial illustration. The repairs-desk scenario below is hypothetical.

The Lovable data exposure report is a reason to check saved work notes and AI summaries separately from a login screen. Crissy Field found internal email summaries belonging to a property manager in records people could access. Before adding confidential writing to a custom app, ask for proof that its database protects every saved copy.[1][3]

On November 15, 2026, the report ranked ninth on Hacker News, with 919 points and 350 comments at the time it was captured. These numbers show how much attention it received, not how many users were affected. The original German coverage came out on November 2. The English report updates that analysis, and PrivacyAffairs covered it on November 14.[1][4][5]

I kept coming back to the email summaries. Even a short message can identify a tenant or reveal a dispute. That matters even if you're not building software. If you're choosing a small work app, find out where it stores the final text.

What the Lovable data exposure figures establish

Crissy Field examined 159,041 projects found through Common Crawl. It then looked more closely at 24,664 sites that returned Supabase records when loaded. Accessing public data this way can be legitimate. Researchers found email address records on 7,403 of those sites, but warned that the 30% figure isn't a confirmed leak rate. The sample also doesn't represent all Lovable apps.[1]

The exposed data included passport details and private chat messages. Researchers couldn't confirm whether anyone had misused it. PrivacyAffairs reported the same limits in its independent coverage.[1][4]

A headline about thousands of leaky apps can lump very different cases together. Letting someone see a restaurant menu in a database isn't a privacy failure. Showing internal correspondence to someone who isn't allowed to see it is. To judge whether access was appropriate, you need to know who the record was meant for.

The report doesn't show that all AI-built apps have this flaw. It can help you ask better questions about a specific app that might store work notes. It doesn't show that every app in this category is safe or unsafe.

A summary needs an access rule of its own

Imagine a repairs desk where an employee writes a private note about a fictional tenant's appointment request. A helper creates a shorter summary for the team dashboard. Later, someone adds search so staff can find past repairs.

There are now several places to check. The source note might have the right restrictions, but a saved summary may follow different rules. Search might also show a snippet. These are examples of things to check, not claims about how any app in the investigation is designed.

Have the builder trace one harmless note through the app. Where is the original stored? Does creating a summary add another database row? Which account can read the note? Include any search result or export that copies part of it. A diagram is fine, but name the actual storage locations.

Supabase uses grants to control which table operations a role can try. Row-level security policies control which rows those operations can access. Signing in gives someone a role, but the policies still have to enforce the intended access rules.[3]

Before changing settings at the repairs desk, decide who can see each piece of information. Only the assigned employee and a supervisor may need access to the full note. The public appointment page should show only fields approved for publication. Let the builder flag designs that don't fit how the team works. Otherwise, a test could pass while enforcing the wrong policy.

Test a saved note with two ordinary accounts

Run this check only in a staging environment you own or have permission to test. Don't include real tenant details. This is a proposed acceptance check, not a test performed on Lovable.

In account A, create an obviously fake note with a distinctive, harmless phrase like "copper umbrella rehearsal," then save its summary. Check that A can find the note again. If the system returns no results for anyone, that doesn't prove legitimate access works.

Next, sign in to account B, which shouldn't be able to access A's note. Ask the developer to check the saved original and summary using the app's usual data routes, including search if available. A blank dashboard doesn't tell you what the database returned.

Test signed-out access on its own. Request only your test records, and access only your own application. Don't search for other people's exposed datasets or copy identifiers from the report.

Record the result in enough detail for someone else to repeat it:

  • The test note, saved summary identifiers, and intended audience.
  • The account and operation you checked, and whether the response included the fake phrase.
  • The application version, test date, and any paths you didn't test, such as exports.

Ask the builder to test write permissions too. The private-note workflow still fails if B can't read A's summary but can replace it. Supabase's guide recommends testing which database actions each user role can and can't perform.[3] If you're choosing software, you don't need to write the tests yourself. Ask the person who maintains the app to show you the results.

Read the security scan's scope before trusting its badge

Lovable's security guide says it runs a Quick scan automatically when you publish. Deep scan is a more thorough code review, but it doesn't run automatically during ordinary work. The guide says automated checks can't guarantee complete security and recommends a professional review for sensitive apps.[2]

A check before publication can catch mistakes. Make sure you know when the version handling your notes was last reviewed. If a builder later adds a summary table or changes sharing, include those changes in the next access review.

Treat past exposure as a separate question. Passing a test today doesn't show who could read records yesterday. If confidential material may already have been accessible, share the evidence with the person who handles security at your organization. Don't post it in a group chat during the investigation.

Choose the destination before you dictate

For a small writing task, you may be able to use a document you already have approval to use. Draft in that document instead of setting up a new notes dashboard. This doesn't establish that the document service is secure, but it avoids adding a storage system no one has reviewed.

DictaFlow can enter dictated text in the app you choose. On desktop, hold a shortcut and speak, then release it to insert text at your cursor.[6] DictaFlow doesn't check or fix a Lovable database or its access rules. Changing how you enter text doesn't change the receiving app's storage permissions.

Before rolling out a custom notes app, ask the builder for a test record showing that account B can't read account A's saved summary. If the app has search, ask them to include the search result too.

Disclosure: Published by DictaFlow with AI assistance from linked sources. The repairs-desk scenario and acceptance checks are hypothetical. We did not reproduce the reported exposure.

Sources

[1] Crissy Field original investigation

[2] Lovable security overview

[3] Supabase row-level security guide

[4] PrivacyAffairs independent coverage, November 14

[5] Hacker News front page, captured November 15

[6] DictaFlow current product facts

More DictaFlow guides