SQL 쿼리 안전하게 노출하기

An MCP server can give a model real data by wrapping a database behind a tool, so the AI asks questions instead of touching raw tables. 🗄️

For safety, start by exposing only read access. Let the model query and report, but keep inserts, updates, and deletes off the table.

Never paste model text straight into SQL. That invites SQL injection, where crafted input rewrites your query and leaks or destroys data.

The fix is parameterized queries: you write placeholders and the driver binds values safely, so input can never change the query shape.

Let the model pass values as typed arguments, then you build the query. The model supplies data, never the SQL text itself.

Keep the SQL fixed in your code and only slot in bound parameters. A locked query shape is the simplest way to stay safe.

Defense in depth: connect with a database read-only user. Even a buggy tool then physically cannot modify or drop your data.

Always cap rows with a LIMIT so one broad query cannot dump a huge table into the model and blow your token budget.

If the model picks a table, validate it against an allowlist. Reject anything not on the list instead of trusting the name.

Format results as clean text or simple records so the model reads them easily. Clarity in, clarity out for the next reasoning step.

Select only the columns the task needs. Leaving out secrets like password hashes keeps private data out of the model entirely.

What protects a SQL tool from injection?

You exposed a database safely: parameterized read-only queries, allowlisted tables, and row limits. Next, wrap a web API. 🎯