Back to Blog
Security7 min readNovember 16, 2026

How Data Breaches Start: Lessons from Exposed API Endpoints

Many data leaks of 2025 and 2026 needed no exploit: an API or a database answered a stranger. Five real cases and six checks for data teams.

The breach that needs no exploit

The usual picture of a data breach is a skilled attacker chaining exploits against a hardened system. The incident reports published in 2025 and 2026 describe something plainer. In many of the leaks that made the news, nothing was broken into. A database listened on a public port with no password. An API returned any record to whoever changed the number in the request. A storage bucket served every file to anyone who had the address.

Data teams now ship exactly this kind of surface: dashboards, internal tools, embedded analytics, small apps on Supabase or Firebase, a FastAPI service in front of the warehouse. Each one is an endpoint on the internet, and each sits next to the most valuable data the company holds. The OWASP Top 10:2025 reflects the trend: Broken Access Control stays in first place, and Security Misconfiguration moved up from fifth to second.

Here are five public cases, what they have in common, and the checks that catch this kind of exposure.

Five incidents, one pattern

DeepSeek: an analytics database with no password (January 2025)

Wiz Research found a ClickHouse database belonging to DeepSeek that was reachable from the internet on two hostnames, with no authentication. ClickHouse accepts SQL over HTTP, so a visitor could run queries from a browser. More than a million log lines were readable, including chat history, API keys and backend details. Wiz reported that it found the database within minutes of looking at the company's public hosts.

Lovable apps: 170 projects with open tables (2025)

A researcher scanned 1,645 apps built with the AI app builder Lovable in March 2025 and found 170 whose Supabase tables answered to the public key shipped in the app's own JavaScript. The finding was filed as CVE-2025-48757. That key is meant to be public. The row-level security policies behind it are the only protection, and on those tables they were missing or too weak.

McHire: a default password and a sequential ID (June 2025)

Researchers Ian Carroll and Sam Curry signed in to a test account on McHire, the recruiting platform McDonald's uses, with the username and password 123456. Once inside, they found an API that returned an applicant's record by numeric ID. Lowering the number returned another applicant's data. The researchers counted about 64 million records behind that ID. Paradox.ai, the vendor, fixed both problems within days and said the researchers had viewed only a handful of records.

Tea: old storage that still answered (July 2025)

Tea, a dating safety app, confirmed that a legacy storage system had been accessed. It counted about 72,000 images, including about 13,000 selfies and photo IDs that users had submitted for verification. Posts about the leak described a publicly readable Firebase storage bucket. The data dated from before February 2024, and the current app no longer used it. It was still online.

Moltbook: a whole database behind a public key (January 2026)

Wiz found that Moltbook, a social network for AI agents, let anyone read and write its Supabase database with the public key from its JavaScript, because row-level security was missing. The exposure covered 1.5 million agent API tokens, about 35,000 email addresses and private messages. Keybrum's write-up of the Moltbook leak explains the mechanism and lists three checks to run on your own app.

None of these needed an exploit. In each case a stranger asked and the system answered.

Why data applications are exposed

  • The endpoint exists to return data. A dashboard API that returns rows is working as designed. Access control is the only difference between returning one customer's rows and returning every customer's rows, and it does not show in a demo.
  • "Internal" rarely stays internal. A Streamlit app, a Superset or Metabase instance, an Airflow UI or a Jupyter server gets a public URL for one stakeholder and keeps it. When Horizon3 checked internet-facing Apache Superset servers in 2023, 2,124 of 3,176 still ran with a default secret key that let anyone forge an administrator session.
  • The browser talks to the database. With Supabase and Firebase, the client holds a public key and the database rules do all the authorization. A table created in a hurry without a policy is readable by every visitor.
  • Old things stay up. The Tea images sat in storage the app had stopped using. Staging copies, a /v1 API and last year's export bucket are still reachable if their rules allow it.
  • Service accounts see too much. The credential behind a dashboard often has read access to the whole warehouse, so one open endpoint exposes far more than the dashboard displays.

Build your architecture diagram now

Datadef generates professional diagrams with AI, 2,000+ cloud icons, column-level data lineage, and an MCP server your coding agent can drive.

Six checks to run this week

  1. List what answers on the internet. Every hostname and port your team has deployed: apps, APIs, BI tools, orchestrator UIs, notebooks, databases. An architecture diagram that marks the internet-facing components makes the forgotten ones visible.
  2. Call every endpoint signed out. Send each API route a request with no token. Any answer other than 401 or 403 needs a written reason.
  3. Call it as the wrong user. Create two accounts. With the second account's session, request the first account's records by ID. OWASP ranks this flaw, broken object level authorization, first in its API Security Top 10.
  4. Read the database rules. On Supabase, list the tables with row-level security off and read every policy whose condition is true. On Firebase, look for allow read, write: if true and for rules that only test request.auth != null. Keybrum publishes free checkers for Supabase policies and Firebase rules that review what you paste.
  5. Request the files nobody should get. /.env, /.git/HEAD, source maps, database dumps, debug routes and the setup page of your BI tool.
  6. Close the database ports. ClickHouse, Postgres, Redis and Elasticsearch should listen on a private interface. Scan your own public IPs to confirm it.

Test from the outside, on every deploy

Every case above was found by someone looking from the outside. In the two Supabase cases, the researchers read the public JavaScript, used the public key and asked the database. A code review misses this kind of exposure, because the settings that matter (database rules, bucket permissions, open ports, hosting options) often live outside the repository.

An external scanner repeats that outside view after each change. Keybrum is a web app security scanner built around it: it loads the deployed app, reads its JavaScript, maps the backend (Supabase, Firebase or a custom API) and runs 338 attack checks, first signed out and then as accounts it creates. Each finding comes with the request that worked and a fix written for a coding agent. It only attacks apps whose owner has proven ownership.

The free check is the place to start: paste a URL, with no account, and see what every visitor's browser already downloads, such as the backend, its keys and the tables named in the code. A full attack run with its report is a one-time €49, and monthly plans from €29 attack again after each deploy.

None of these leaks took an exploit to find, and the checks that catch them take an afternoon. Run them on your own apps first.