Before your team uses it: a security checklist for an app you built with AI
Before colleagues use an app you built with AI, check ten things: every page asks for sign-in; only the right people can sign in; no passwords or keys sit in the code; you know every outside service the app sends data to; its packages are known; typed text can't run as a command; the data is backed up and you have tried a restore; you know where it is stored; someone else can take it over; and you can see what happened. Each check takes minutes, and you can run it on any host.
Sources read on 13 November 2026. A plain list of the ten checks to print is at the end.
Most of these you can do by asking your assistant, because it can read the code. Treat what it says as a lead to confirm, not as proof. Where a check has a documented example, it is listed, with its date and source. The standards named are the OWASP Top 10 (2021 edition).
1. Does every page ask for sign-in?
What goes wrong. OWASP lists "force browsing to authenticated pages as an unauthenticated user" as a common failure of access control (A01). In May 2025 a vulnerability record, CVE-2025-48757, described database rules on apps built with Lovable that let anyone read or write tables without signing in. The researcher who reported it scanned 1,645 Lovable-built apps and found 170 with the problem. Lovable disputes the record and says each customer is responsible for protecting their application's data (NVD, the researcher's statement).
Check it. Open the app in a private window, so you aren't signed in, and paste its address. Then try a few deeper addresses you know, such as a record's page you copied while signed in. Each should send you to sign-in. You can also ask your assistant to list every route and endpoint and say which work without sign-in. Good looks like a sign-in screen every time, with data endpoints included.
2. Can only the right people sign in?
What goes wrong. In July 2025 Wiz Research reported that on Base44, a platform for building apps with AI, giving only an app's non-secret ID to undocumented sign-up endpoints could create a verified account on a private app. Wiz says it got past the platform's sign-in controls, including single sign-on, and that the flaw was fixed in under 24 hours with no evidence of past abuse (Wiz). OWASP also lists weak or default passwords and missing multi-factor sign-in (A07).
Check it. Open the app's sign-in settings and see who is allowed: named people, a company email domain, or anyone. Try signing up with a personal email address and see whether it gets in. Remove people who have left. Good looks like a short, named list or one company domain.
3. Are there keys or passwords in the code?
What goes wrong. OWASP's secrets guidance says keys, database credentials and certificates are often hardcoded "in plaintext, littered throughout configuration files", and that a secret that may be compromised must be revoked (OWASP). GitGuardian, which sells secret scanning, counted 28.65 million new hardcoded secrets in public GitHub commits in 2025 (its 2026 report). A public Supabase "anon" key in front-end code is designed to be public. The risks are a service key, or missing row-level rules.
Check it. Ask: "Search every file in this project, including config and .env files and the git history, for API keys, passwords, tokens and connection strings, and list where each one is." Move anything found into the host's secret settings and replace the key at the provider, because a key that sat in code should be treated as seen. Steps for that are in keeping API keys out of the code Claude writes. Good looks like an empty list.
4. Do you know every outside service it talks to?
What goes wrong. OWASP's misconfiguration entry lists "unnecessary features enabled or installed" (A05). For personal data there is a legal side too: the UK's data protection regulator says people risk losing the protection of UK data protection law if their information is sent outside the UK, and treats sending it to a separate organization abroad as a "restricted transfer" (ICO).
Check it. Ask: "List every outside web address, API and service this app calls or loads, including scripts and fonts, what data goes to each, and where in the code." Confirm each is something you meant to use and that your company allows that data to go there. Good looks like a short list you can explain.
5. Are its packages known and current?
What goes wrong. OWASP says you are likely exposed if you don't know the versions of the components you use, nested ones included (A06). In September 2025 CISA warned that a self-replicating worm had compromised over 500 packages on npm and advised pinning versions to releases from before 16 September (CISA). Assistants also sometimes name packages that don't exist: one academic study of 576,000 code samples found at least 5.2% of packages suggested by commercial models were made up, and 21.7% for open-source models (Spracklen and others).
Check it. Ask: "List every package this app uses with its version, run the package manager's audit (npm audit or pip-audit), and tell me what is flagged as vulnerable or not found on the registry." Make sure a lockfile is committed. Good looks like a list with nothing flagged, or flags you understand.
6. Can typed text run as a command or a database query?
What goes wrong. An app is open to injection when user-supplied text is "not validated, filtered, or sanitized" and queries are built by gluing text together. OWASP's preferred fix is a safe interface that takes the values separately, called a parameterized query (A03).
Check it. Ask: "Find every place where text from a form, a URL or an upload is put into a SQL query, a shell command, eval or an HTML template. Is each one parameterized or escaped? Show me the lines." Good looks like every one of them using the safe form.
7. Is the data backed up, and have you tried a restore?
What goes wrong. In July 2025 SaaStr's founder Jason Lemkin reported that Replit's AI agent had deleted a production database. Replit's chief executive called it unacceptable, said Replit keeps backups with one-click restore of a project's state, and the company said it would separate development and production databases (The Register). The UK's National Cyber Security Centre says that once you have a backup you should know how to restore it and check it holds your important data (NCSC).
Check it. Find when the last backup ran and how long backups are kept. Restore one into a copy, or ask the host to, check that your latest records are in it and note how long it took. Good looks like a restore you have done once.
8. Do you know where the data is stored?
What to know. The app, its database, its file storage and its backups can each be in a different place. In the host's settings find the region of each, and list any outside service from check 4 that receives the data. Compare that with your company's rules for customer or staff data. Good looks like one answer per place, written down.
9. Can someone else take it over?
What to know. An app that only one person can change stops being a team tool the day that person is away or leaves. Add a second owner or admin on the host. Keep the code in a company-owned repository, not on one laptop. Write down where the keys, the domain and the billing live. Good looks like a second person who can deploy a change and restore a backup without you.
10. Can you see what happened?
What goes wrong. OWASP lists logins and failed logins that aren't logged, and logs kept only locally, as failures (A09).
Check it. Find where the host shows sign-ins and app logs. Check that you can see who signed in yesterday and what the app printed when something failed, and for how long logs are kept. Good looks like an answer to "who opened it on Tuesday?".
What Haechy does by default
Details are on the security page, in approving outside services, backups and restore and safe changes, and for IT in the page for IT admins. The checks apply to any host. If you need to share the list with a manager, the printable version is below, and sharing an app you built with Claude covers the question of who gets access.
The ten checks, to print
Before colleagues use the app
[ ] 1. Every page asks for sign-in (tried in a private window)
[ ] 2. Only the right people can sign in (no old colleagues, no open sign-up)
[ ] 3. No keys or passwords in the code, and any that were are replaced
[ ] 4. I have a list of every outside service the app calls
[ ] 5. I know which packages it uses, and the audit shows nothing flagged
[ ] 6. Text typed into forms is never run as a command or a database query
[ ] 7. The data is backed up, and I have restored one
[ ] 8. I know where the data is stored, and that is allowed
[ ] 9. A second person can deploy a change and restore a backup
[ ] 10. I can see who signed in and what the app printed when it failed Haechy is an independent product of DT Global Ventures Ltd. It is not made, sponsored or endorsed by Anthropic. Claude is a trademark of Anthropic, PBC.