Story

Is Lovable safe? Supabase data exposure in customer apps

Some Lovable apps finish loading for a visitor who never logged in, and the page has already been sent rows from a Supabase database. On 14 November 2026 (UTC) the research behind that picture reached the Hacker News front page, at about 387 points. The question on the thread is a plain one: is Lovable safe?

The English report is from Crissy Field GmbH, a security firm in Hamburg. The work was done with Handelsblatt. Lina Knees reported the findings on 2 November 2026, in German. What the firm describes is misconfigured access rules in customer apps. A breach of Lovable is not the finding.

A short timeline

The public record is a few weeks long. Supabase exposure was already a story before anyone published a Lovable count.

When What is on the record
25 September 2026 TechCrunch reports UpGuard's Supabase exposure findings.
14 October 2026 UpGuard publishes its study: 16,326 databases with readable tables.
30 October 2026 Supabase stops auto-exposing new public tables on existing projects. Existing grants stay. Row level security is unchanged.
2 November 2026 Handelsblatt publishes Crissy Field's findings on Lovable apps.
3 November 2026 India Today, citing Handelsblatt: Lovable says the problem is industry-wide, and that customers are responsible for their apps and data.
4 November 2026 Cybernews describes missing or unconfigured row level security, not a hack of Lovable.
10 November 2026 Lovable publishes a security post on scanners, a Fix action, and auto-fix.
14 November 2026 The Crissy Field report reaches the Hacker News front page. The author answers. Ethics and liability get argued.

UpGuard's study is its 14 October post. The 25 September coverage is TechCrunch. The grant change is the Supabase changelog. The Lovable findings, including the Handelsblatt date, are on Crissy Field's page. The 3 November coverage is India Today. The 4 November piece is Cybernews. Lovable's 10 November post is Find it, fix it, ship it. The thread is Hacker News.

Is Lovable safe? What Crissy Field counted

Crissy Field used Common Crawl. From December 2025 through September 2026 it counted 159,041 Lovable projects. Of those, 34,545 used Supabase. The firm says this sample is not representative of all Lovable apps. Lovable’s own scale, reported later by Cybernews, is a different number, and it is Lovable’s: more than 50 million projects, and 1.2 million new ones a week. The crawl is a slice. The company figure is a company figure.

Crissy Field also counted 24,664 sites that sent Supabase records when the page loaded, with no login. Crissy Field says that is not a problem by itself. The data may be meant to be public.

The next count is the one that spread. With further read-only requests, records containing email addresses were found in 10,560 tables, across 7,403 of those 24,664 sites. That is the 30 percent. Crissy Field writes that the 30 percent figure “should therefore not be read as a confirmed data leak rate”. Read it as what this sample returned, under the method the firm used, on sites already sending Supabase records to a logged-out page. It is not a measured leak rate for every Lovable app.

The examples in the report are specific. The apps are not named. A German lifestyle forum exposed more than 80,000 profiles and more than 200,000 chat messages. Several event organisers and restaurants exposed tens of thousands of ticket and food-sale records, including emails, phone numbers, postal addresses, payment information, and ticket codes. A property management company exposed AI-generated summaries of internal emails. A service provider exposed more than 1,000 records with passport numbers, expiry dates, and nationalities. In one case, 280,000 emails sent to a provider’s customers were accessible.

Crissy Field could not establish whether any data was misused. It did not publish the access method. It did not name the apps. This story does not describe that method either. A builder checking their own project does not need it.

The firm’s line on the platform is a full sentence, and it is short enough to quote whole: “An easy-to-use development platform does not remove the need for an expert review of access permissions before an application handles confidential data.”

How much a researcher should pull was its own argument, once the report hit Hacker News. jacquesm wrote: “POC -> fine Actual mass exfiltration -> not fine.” neathack, the author of the Crissy Field post, replied that they retrieved only enough to judge whether the data was privacy-sensitive, and never more than the first page.

Did Lovable get hacked?

No. The published research describes misconfigured access rules in customer apps: Supabase row level security that was missing or left too open. It does not describe a breach of Lovable.

neathack wrote on the thread: “I can confirm that no advanced hacking techniques were used”. neathack also wrote that an earlier submission had gone quiet: “I submitted it to HN two weeks ago, but it didn’t gain much traction”. The dataset, neathack wrote, was shared with Lovable “so they could contact their customers and have the database access issues resolved”.

Cybernews, on 4 November 2026, described the cause the same way: missing or unconfigured row level security, not a hack of Lovable itself. The scale figures in that piece are Lovable’s, as Cybernews reports them: more than 50 million projects, and 1.2 million new ones a week. What payment-verified AI-built apps make is a separate record from how those apps store other people’s rows.

India Today, on 3 November, citing Handelsblatt, reported Lovable’s response. Lovable said the problem was not limited to its platform and was widespread across the industry. It said its terms place responsibility for applications and data handling on customers.

A week later Lovable published its own security post. Find it, fix it, ship it, dated 10 November 2026, is by Igor, Head of Security. The post says scanners check code, logic, database schemas, and row-level security policies on every publish. It describes a free deep scan, a Fix action, an “Auto-fix security issues” project setting, and publishing controls that can block critical findings. Scheduled scans on paid plans, the post said, were coming this month. Igor’s line on the failure is: “The most common issue we see in Lovable apps is data access rules that are too open”.

The same class of open database shows up far outside Lovable. UpGuard, on 14 October 2026, wrote up about 300,000 domains with Supabase indicators. It found 16,326 databases with readable tables. Over half had indicators of personal data. TechCrunch had reported UpGuard’s findings on 25 September, before that full study. Readable Supabase tables, in UpGuard’s count, are a wider fact than Lovable.

What is Supabase row level security?

Row level security is the Postgres control on a table: which rows a request may read or change. Supabase documents that control in its guide to securing your API and its row level security guide.

The default is not the same for every table. The securing-the-API guide says tables created in the Dashboard have row level security enabled by default. Tables created in the SQL Editor, or through other tools, need row level security enabled explicitly. A table that came from the Dashboard is in the first group. A table that came from the SQL Editor or from another tool stays in the second group until someone enables the control.

On 30 October 2026 Supabase changed a different setting from row level security. From that date, on existing projects, new tables in the public schema are not exposed to the Data API without an explicit grant. Existing tables keep their grants. The changelog says “RLS behavior remains unchanged.” A new table that nobody granted is harder to reach through the API. A table that already had a grant keeps it. Row level security on either table is the setting the changelog says it did not change.

Supabase’s 2025 security retrospective describes the work that came before that October change: row level security by default for tables made in the dashboard, an event trigger to auto-enable row level security, security advisors, and the split between publishable keys and secret keys. The retrospective is the company’s account of those controls. The docs still tell you to enable row level security yourself on tables that did not get it from the Dashboard.

On the Hacker News thread, two comments put Supabase next to the apps. SOLAR_FIELDS wrote: “The part people are ignoring is Supabase’s culpability in this matter”. YuechenLi wrote: “One of the golden rule of webdev is ‘never build your own auth’ for a reason.” Those are comments on the thread. They are not a finding from Crissy Field, and they are not Supabase’s position.

Who is responsible for security in a Lovable app?

Crissy Field notes two messages at once. Lovable’s terms and conditions place responsibility for applications and data handling on customers. The marketing stresses that no technical knowledge is needed. India Today’s 3 November account, citing Handelsblatt, reports Lovable saying the same thing about the terms, and saying the problem was widespread across the industry.

The 10 November post is the platform adding checks on the path between a project and a public URL: scanners on every publish, a free deep scan, a Fix action, auto-fix, and a way to block a publish on critical findings. The terms, as Crissy Field and India Today record them, still leave the application and the data with the customer. Both are on the record.

Supabase’s piece is the defaults, the advisors in the 2025 retrospective, and the 30 October grant change for new tables. Existing tables keep the grants they had. The changelog says “RLS behavior remains unchanged.” A database company can change the default for a table created tomorrow and leave a table created yesterday as the customer configured it.

The thread split on who should carry the cost. Oras wrote: “They are an easy target, they don’t need any sophisticated hacking knowledge”. hn_throwaway_99 wrote: “the only thing that can fix this are real, stiff liability consequences”.

The speed that puts these apps online is already a separate argument. Shipping fast with AI is about a first version getting cheap, and about the permission check that a careful build still has to catch. A product that stores someone else’s data is the case that page treats as not disposable. Whether vibe coding is killing indie hacking is the claim that the hard part stopped being the build. More people can ship. The rows in the database are still someone’s email address, phone number, or passport number.

A different week produced a different security story about a shipped product. Reports that Humble Bundle’s two-factor check vanished in a redesign are about what a live site required at sign-in, and about a vibe-coding argument that those reports do not settle. The Lovable research is about what a logged-out page was sent from the database.

How do I check if my Lovable app is exposing data?

The checks below are the ones Supabase and Lovable document, plus the logged-out look that matches what Crissy Field counted. They are a review of your own project. They are not a method for reading someone else’s.

  1. Turn on row level security for every table the Data API exposes. Supabase’s row level security guide is the reference, and the securing-the-API guide is the one that ties it to the API. Cybernews described the exposure as missing or unconfigured row level security. Lovable called the common failure data access rules that are too open.

  2. Check where the table was created. Tables created in the Dashboard have row level security enabled by default. Tables created in the SQL Editor, or through other tools, need it enabled explicitly. That split is Supabase’s, in the securing-the-API guide. Look at the tables that did not come from the Dashboard.

  3. Read the 30 October change as a change about grants, then check the old tables anyway. New tables in the public schema, on existing projects, are not exposed to the Data API without an explicit grant. Existing tables keep their grants. “RLS behavior remains unchanged.” If a table was reachable last month, the changelog does not say it became unreachable on 30 October.

  4. Act on Lovable’s scanner. The 10 November post says the scanners look at code, logic, database schemas, and row-level security policies on every publish. Use the free deep scan. Use the Fix action. The project setting named in the post is “Auto-fix security issues”. Publishing controls can block critical findings. The post said scheduled scans on paid plans were coming this month, meaning the month of that post.

  5. The publishable key in the browser is not a secret. Supabase’s 2025 retrospective covers publishable keys and secret keys. The exposure Crissy Field described is rows coming back to a logged-out page. That is the access rules on the tables. The presence of the publishable key is a separate fact.

  6. Open the published app while logged out. Crissy Field’s 24,664 sites are sites that sent Supabase records on page load, with no login. The firm says that alone is not a problem, because the data may be meant to be public. Look at your own pages the same way: a private window, no session. If the page shows email addresses, phone numbers, postal addresses, payment information, ticket codes, or passport details, those are the kinds of records the study flagged. If it shows only records you meant to publish, that is the case the firm said was not a problem by itself. This look is not the expert review in Crissy Field’s sentence.

  7. Do the review again after a large change. The scanners run on every publish, which is the moment a large change becomes the live app. Crissy Field’s sentence still sits underneath the scanner: “An easy-to-use development platform does not remove the need for an expert review of access permissions before an application handles confidential data.”

One neighbouring chore is easy to mix up with this one. Rotating API keys after the Trigger.dev and GitBook breaches is the job when a disclosure says a vendor’s systems may have been read. The Lovable research describes access rules on customer apps. Cybernews described that research’s cause as not a hack of Lovable. Rotating a key is one job. Enabling row level security is the other. Each leaves the remaining job still to do.

Crissy Field counted a real sample, said the 30 percent is not a confirmed leak rate, and said it could not show misuse. Lovable called overly open data rules the common failure, and shipped scanners on 10 November. Supabase had already documented the defaults, and on 30 October it stopped handing new public tables to the Data API without a grant. UpGuard’s October count is the reminder that readable Supabase tables were already a wider fact.

What do you think? When anyone can ship an app that stores other people’s data, who should own security: the builder, the platform, or the database?

FAQ

Is Lovable safe?

Crissy Field GmbH, a Hamburg security firm working with Handelsblatt, describes misconfigured access rules in customer apps built on Lovable and Supabase. It does not describe a breach of Lovable. Using Common Crawl, the firm counted 159,041 Lovable projects from December 2025 through September 2026, of which 34,545 used Supabase. It found 24,664 sites that sent Supabase records on page load with no login, and said that alone is not a problem because the data may be meant to be public. Records containing email addresses were then found in 10,560 tables across 7,403 of those 24,664 sites, which is 30 percent. Crissy Field writes that this 30 percent figure should not be read as a confirmed data leak rate, and that the Common Crawl sample is not representative of all Lovable apps. The firm could not establish whether any data was misused, and it did not publish the access method or name the apps. Lovable told coverage on 3 November 2026 that the problem was widespread and that its terms put applications and data handling on customers. On 10 November, Igor, Lovable's Head of Security, wrote that the most common issue in Lovable apps is data access rules that are too open, and described scanners that run on every publish.

Did Lovable get hacked?

No. Researchers describe misconfigured access rules in customer apps, missing or unconfigured Supabase row level security, not a breach of Lovable. On the Hacker News thread of 14 November 2026, neathack, the author of the Crissy Field post, wrote that no advanced hacking techniques were used. Cybernews, on 4 November 2026, described the same cause and reported Lovable's own scale figures of more than 50 million projects and 1.2 million new ones a week. India Today, on 3 November, citing Handelsblatt, reported that Lovable said the problem was not limited to its platform and was widespread across the industry. neathack also wrote that the dataset was shared with Lovable so the company could contact customers and have the database access issues resolved. Crissy Field could not establish whether any data was misused. A wider UpGuard study, published 14 October 2026, found readable tables in Supabase databases across the web, so the pattern is not unique to Lovable.

What is Supabase row level security?

Row level security is the Postgres control Supabase documents for deciding which rows a client can read or change through the API. Supabase's securing-the-API guide says tables created in the Dashboard have row level security enabled by default, and tables created in the SQL Editor or through other tools need it enabled explicitly. A changelog entry applied to existing projects on 30 October 2026 says new tables in the public schema are not exposed to the Data API without an explicit grant, existing tables keep their grants, and row level security behavior remains unchanged. That grant change does not replace row level security on tables already exposed. Lovable's 10 November 2026 security post says the most common issue in Lovable apps is data access rules that are too open. Supabase's 2025 security retrospective covers row level security by default for dashboard tables, an event trigger to auto-enable it, security advisors, and the split between publishable keys and secret keys.

How do I check if my Lovable app is exposing data?

Enable row level security on every table exposed through Supabase's Data API. Tables made in the Dashboard have it on by default. Tables made in the SQL Editor or through other tools do not, unless you enable it. The 30 October 2026 grant change applies to new tables in the public schema. Existing tables keep their grants, and row level security behavior is unchanged, so check tables that were already there. In Lovable, read the security scan that runs on every publish. The 10 November post describes a free deep scan, a Fix action, an Auto-fix security issues setting, and publishing controls that block critical findings. Scheduled scans on paid plans were described as coming that month. The publishable key in the browser is not a secret. Open the published app while logged out and see whether records you treat as private are on the page. Crissy Field says data sent on load can be data you meant to publish. Review the rules again after a large change. An easy-to-use platform does not remove the need for an expert review before an app handles confidential data.

Who is responsible for security in a Lovable app?

Crissy Field notes that Lovable's terms place responsibility for applications and data handling on customers, while the marketing stresses that no technical knowledge is needed. India Today, citing Handelsblatt, reported the same terms position from Lovable, and reported that Lovable called the problem widespread across the industry. On 10 November 2026 Lovable published scanners, a Fix action, auto-fix, and publish blocks for critical findings, which is platform work sitting next to those terms. Supabase documents the row level security defaults, and from 30 October 2026 it stopped exposing brand-new public tables to the Data API without a grant. On Hacker News, hn_throwaway_99 wrote that only stiff liability consequences can fix this, SOLAR_FIELDS pointed at Supabase, and YuechenLi cited the rule against building your own auth. The public record shows three parties publishing different pieces. It does not pick one owner.