Why a web security list concerns data engineers
The OWASP Top 10 is the reference ranking of web application security risks. The 2025 edition, the eighth, was built from data on more than 2.8 million applications. It is written for web developers, and data engineers tend to skip it.
That made sense when data work ended at the warehouse. Today a data engineer ships a dashboard, an embedded analytics page, a FastAPI service over the warehouse, a Streamlit app or a small tool on Supabase. Each of these is a web application, and it sits closer to the data than most of the product does.
Three changes from the 2021 edition are worth knowing. Security Misconfiguration moved from fifth place to second. Server-side request forgery (SSRF) is now part of Broken Access Control. Two categories are new: Software Supply Chain Failures and Mishandling of Exceptional Conditions.
Three categories to check first
A01 Broken Access Control: dashboards that show everyone's rows
In a data application, access control is mostly a question of rows: which tenant, which region, which customer. The usual failures look like this:
- A dashboard API takes
?customer_id=42and filters on it without checking that the caller belongs to customer 42. Change the number and another customer's revenue comes back. - An embedded dashboard applies the tenant filter in the browser, or signs an embed token whose tenant parameter the client can still change.
- Row-level security is defined in the BI tool and not in the warehouse, so the API and the notebook that query the same tables skip it.
- A Supabase policy only checks that the caller is signed in. Where sign-up is open, anyone can create an account.
- A CSV export route was added after its page and never received the page's permission check.
- A "test connection" or "import from URL" feature fetches any address the user types, including internal hosts and the cloud metadata service at
169.254.169.254. This is SSRF, and data tools are full of features shaped like it.
The test needs two accounts in two tenants. Signed in as the second, request the first one's dashboards, exports and embed URLs. Every answer should be a 403, a 404 or an empty result.
A02 Security Misconfiguration: the config is public, the rules are the lock
Supabase and Firebase put a config object in the browser: a project URL and a public key. That is by design and it is not a leak. What matters is what the key can reach: a table with row-level security off, Firebase rules left at allow read, write: if true, a public storage bucket. The credentials that must never ship to the browser are the Supabase service_role key and Firebase admin credentials, because they bypass every rule.
Self-hosted data tools add their own defaults. In 2023, Horizon3 found that 2,124 of 3,176 internet-facing Apache Superset servers ran with a default secret key, which lets anyone forge an administrator session (CVE-2023-27524). The same year, Metabase fixed a flaw that let an unauthenticated visitor run commands on the server through a setup token that stayed readable after installation (CVE-2023-38646).
The rest of the category is familiar: an orchestrator or notebook UI without a login, a database port open to the internet, a served .env file, debug mode in production, missing security headers. For the last one, Keybrum's free security headers checker grades the headers your live app sends and gives the line to add for each missing one.
A05 Injection: SQL built from a filter
Analytics code builds SQL dynamically more than any other kind of code. The table, the columns, the grouping and the sort order all come from what the user picked. Bind parameters protect values, and they cannot stand in for identifiers. So the ORDER BY column, the GROUP BY dimension and the table name end up formatted into the string, in an f-string, a Jinja template or a template literal.
- Pass every value as a bind parameter.
- Check every identifier (column, table, sort direction) against an allowlist. Do not escape them by hand.
- Give the application a read-only database role, limited to the views it needs, with a statement timeout. A successful injection then has a ceiling.
- Treat SQL written by a language model as untrusted input, and run it under the caller's own permissions.
Mature tools get this wrong too. In 2024, Grafana disclosed CVE-2024-9264: its experimental SQL Expressions feature passed user queries to DuckDB without sanitizing them, so a user with Viewer permission could read files and run commands on a server where the DuckDB binary was installed.
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.
The other seven, in data terms
- A03 Software Supply Chain Failures. A data stack is hundreds of Python packages, dbt packages and Docker base images, and the pipelines that install them hold warehouse credentials. In March 2026, two malicious releases of LiteLLM stayed on PyPI for about 40 minutes and carried a credential stealer. Pin versions with a lockfile and hashes.
- A04 Cryptographic Failures. Database connections with
sslmode=disable, embed tokens signed with a short or default secret, exports with personal data left unencrypted in object storage. - A06 Insecure Design. An export route with no row limit. A warehouse or AI query route with no rate limit, where one caller can run up the compute bill. A request that carries its own
roleortenant_idand a server that believes it. - A07 Authentication Failures. Shared BI logins, service account keys pasted in notebooks, warehouse users without multi-factor authentication. In the 2024 attacks on Snowflake customers, about 165 organizations were notified that they were potentially exposed. The attackers used stolen passwords on accounts that had no MFA.
- A08 Software or Data Integrity Failures. Loading a pickle file, such as a model or a cached dataframe, executes code. If other people can write to the bucket it comes from, they can run code in your pipeline.
- A09 Security Logging and Alerting Failures. The question to answer is who queried this table, when, and through which application. Keep the warehouse query history and the API access logs, and alert on unusual export volumes.
- A10 Mishandling of Exceptional Conditions. An API error that returns the failed SQL statement, a stack trace or a connection string. An authorization check that lets the request through when the identity provider times out.
What you can test from outside, and what you cannot
Five of the ten categories show plainly on the deployed application: A01, A02, A05, A07 and A08. Four show in part: A03, A04, A06 and A10. A09 does not show at all, because nobody outside can see your logs. Whatever shows from outside can be tested the way a stranger would attack it.
An automated penetration test of the live app covers that part. Keybrum reads the application's JavaScript, maps the backend, and runs 338 attack checks, first signed out and then as accounts it creates. It covers Supabase tables, functions and storage, Firebase rules, custom API routes including GraphQL and tRPC, leaked keys, exposed files, and SQL injection confirmed by a database error. Every finding comes with the request that worked. Keybrum also publishes a table of what it attacks in each OWASP category and what it leaves out.
You can start with a free preview of what your app exposes: it takes a URL and no account, and shows 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 the monthly plans attack again on each deploy.
A one-hour OWASP pass for a data application
- A01. With two accounts in two tenants, request the other tenant's dashboards, exports and embed URLs.
- A01, A07. Signed out, call each API route and open each tool URL: BI, orchestrator, notebooks.
- A02. On Supabase, list the tables with row-level security off. On Firebase, read the rules for
if true. Request/.envand/.git/HEADon your domain. - A07. Search the built JavaScript for
service_role,sk-and-----BEGIN. - A05. Search the code for SQL assembled with f-strings,
.format()or template literals. Check that every dynamic identifier has an allowlist and that the database role is read-only. - A10. Send an API route a body it cannot parse and read the error that comes back.
- A03. Run
pip-auditornpm audit, and confirm installs come from a lockfile. - A09. Confirm you could tell who read a given table last week.
One prerequisite sits under all eight steps: knowing what exists. Access control cannot be checked on a service nobody remembers deploying. An architecture diagram that marks the internet-facing components and the stores that hold personal data gives the review its scope. Datadef generates that diagram from your repository and keeps it in step as the platform changes.