Can apps built with AI be hacked?
Yes, and the reason is consistent across tools. AI coding tools are very good at making a feature work and less focused on the settings that keep it safe. So the app runs perfectly and quietly skips things like database access rules, hidden keys, and security headers.
That means most real incidents are not sophisticated attacks. They are ordinary requests that succeed because a protection was never turned on: reading a table that has no access rules, copying a key that shipped to the browser, or downloading source that should not be public.
Because these are configuration gaps rather than deep flaws, they are both easy to find and easy to fix. A read-only scan of the live app shows which ones are present, and each comes with a plain-English fix.
Scan your AI-built app
What it means
Anyone can read this table without logging in. That can be deliberate, such as reference data or a public directory, or a missing access policy; only you can say which. xlogs read one row's column names and kept none of its values.
Why AI tools cause it
Supabase only limits who can read a table once you turn on Row Level Security (RLS) and write access rules for it. An AI tool can build a working app without doing that, and a table without RLS, or with a policy that lets everyone read, answers anonymous requests.
How xlogs checks it
xlogs finds your public Supabase URL and anon key in your app (the same ones your frontend uses), asks the database for its table list, reads one row from each listed table as an anonymous user, and flags any that return rows. Hosted Supabase refuses that list to public keys, so on a hosted project the result says the tables could not be checked. Read-only: it never writes or deletes.
How to fix it
The goal: Row Level Security is enabled on every table, and each table returns rows only to the users it is meant for: a private table answers no anonymous request, and a table that is public on purpose has a read policy that says so.
- Decide whether the affected spot is meant to be readable without a login
- If it is not, enable Row Level Security (RLS) on the affected spot and add a policy so only the users allowed to see a row can read it
- If it is, keep it public on purpose, and record that decision so a review does not reopen it
- Do not switch on a restrictive policy for a table your public pages read without a replacement read policy, or those pages stop working
- Re-check that an anonymous request returns what you decided it should
In my app (the spot the scan shows): anyone can read this table; confirm that is deliberate. Please make this true: Row Level Security is enabled on every table, and each table returns rows only to the users it is meant for: a private table answers no anonymous request, and a table that is public on purpose has a read policy that says so. Steps: Decide whether the spot the scan shows is meant to be readable without a login; If it is not, enable Row Level Security (RLS) on the spot the scan shows and add a policy so only the users allowed to see a row can read it; If it is, keep it public on purpose, and record that decision so a review does not reopen it; Do not switch on a restrictive policy for a table your public pages read without a replacement read policy, or those pages stop working; Re-check that an anonymous request returns what you decided it should. Then tell me exactly what you changed, and do not print any secret values back to me or write one into a commit message, a pull request description or a log.
Then verify: After you decide and deploy, xlogs repeats the same anonymous read of each table the database lists and reports what comes back: a table you made private should no longer return rows, and one you keep public still will. If the table list is refused, it says the tables could not be checked.
Full step-by-step fix guide, with a copy-paste block for each AI tool →
Common questions
Which AI coding tool is the most secure?
The differences are small, because the gaps come from the same skipped steps regardless of tool: access rules, key handling, and headers. What matters more is checking the app you actually shipped, since any of these tools can produce an app that works and still exposes data.