Safe SQL Execution

Safe SQL execution

Supabase Studio executes SQL statements directly against the user's database. Because this is the authenticated user's own database, our security model is different from most frontend applications: a user should be able to execute any SQL statement, as long as it is proven that they themselves authored it. What we SHOULD NOT ALLOW is execution of SQL statements that can be influenced by an attacker, such as through URL parameters.

Security model

The security model for SQL execution in Supabase Studio is based on the principle of "proven authorship". This means that a user should only be able to execute SQL statements that they have explicitly authored, and not statements that can be influenced by external input.

There are three classes of SQL fragments:

1. Hardcoded within the application code. These are safe to execute because they cannot be influenced by an attacker. They can be marked with the safeSql utility with pg-meta:

   import { safeSql } from '@supabase/pg-meta'
   
   const sql = safeSql`
     SELECT *
     FROM users
     WHERE id = 1
   `

   safeSql automatically creates a string of the branded type SafeSqlFragment. (See Provenance Tracking below.)

2. Third-party influenceable. These are SQL fragments that can be influenced by an attacker, such as through URL parameters or LLM output. These should be marked with the untrustedSql utility with pg-meta:

   import { untrustedSql } from '@supabase/pg-meta'
   
   const unsafeQuery = searchParams.get('query')
   const querySql = untrustedSql(unsafeQuery)

   untrustedSql creates a string of the branded type UntrustedSqlFragment. (See Provenance Tracking below.)

3. User-authored. These are SQL fragments that are authored by the user themselves within the UI, for example in a text input field. Because the user is the author, these should be considered safe to execute.

   However, there is a caveat, where third-party and user-authored code can mix, contaminating the user-authored code (for example, if an input is prefilled from an unsanitized URL parameter). Provenance tracking helps us track these cases.

   For example, a safe input component could be implemented as follows by requiring that its placeholder and controlled value are of type SafeSqlFragment. In this case we can use its onChange to promote the user input to SafeSqlFragment type, because we know that the user is the author of the input. An implementation of this is in @apps/studio/components/ui/SafeSqlInput.tsx:

   import { rawSql, type SafeSqlFragment } from '@supabase/pg-meta'
   import type { ChangeEvent, ComponentProps } from 'react'
   import { Input } from 'ui-patterns/DataInputs/Input'
   
   type InputProps = ComponentProps<typeof Input>
   
   export type SafeSqlInputProps = Omit<
     InputProps, 'placeholder' | 'value' | 'onChange'
   > & {
     placeholder?: SafeSqlFragment
     value: SafeSqlFragment
     onChange?:
       (event: ChangeEvent<HTMLInputElement>, value: SafeSqlFragment) => void
   }
   
   export const SafeSqlInput = ({ onChange, ...props }: SafeSqlInputProps) => (
     <Input
       {...props}
       onChange={(event) => onChange?.(event, rawSql(event.target.value))}
     />
   )

   This is pretty much the ONLY VALID USE CASE of the rawSql export from pg-meta, and it should be used with caution.

Provenance tracking

Branded types are used to track the provenance of SQL fragments. The types, exported from pg-meta, are:

• SafeSqlFragment: represents SQL fragments that are safe to execute, because they are either hardcoded in the application or authored by the user themselves.
• UntrustedSqlFragment: represents SQL fragments that can be influenced by an attacker, such as through URL parameters or LLM output.

These are valid ways to generate a SafeSqlFragment:

• Using the safeSql utility from pg-meta to create hardcoded SQL fragments.
• Using the sanitization utilities from pg-meta to sanitize untrusted input and promote it to a SafeSqlFragment:
  • ident
  • literal
  • keyword
• Using the safe SQL manipulation utilities:
  • joinSqlFragments (from pg-meta)
  • trimSafeSqlFragment (from apps/studio/lib/sql.ts)

UntrustedSqlFragments can be generated from raw strings using untrustedSql().

There is also a union type, DisplayableSqlFragment, which represents SQL fragments that can be safely displayed in the UI, but not necessarily executed. This includes both SafeSqlFragment and UntrustedSqlFragment.

Security of SQL round-tripped from the user's database

SQL derived directly from catalog tables (e.g., function definitions, RLS expressions, etc.) is considered safe, and it is promoted AT THE POINT OF BEING QUERIED from the database. In most cases, this is in an apps/studio/data/*/.ts file, in the utility function that makes the API or database fetch.

A critical exception to the safety of SQL round-tripped from the database is user snippets. These must NEVER BE CONSIDERED SAFE because they are both (a) externally influenceable and (b) auto-saved. The snippet type uses the unchecked_sql property, which is an UntrustedSqlFragment, to enforce this.

Promoting SQL fragments to SafeSqlFragment type

Given an insecure string or UntrustedSqlFragment, how do we promote it safely to a SafeSqlFragment?

Sanitization utilities

This is the preferred method when the input is sanitizable, e.g., it is a relation name, a column name, will be compared as a literal, etc.

The pg-meta library provides the following sanitization utilities that can be used to safely promote untrusted input to SafeSqlFragment:

• ident: for sanitizing identifiers such as table names or column names.
• literal: for sanitizing literal values that will be used in SQL statements.
• keyword: for sanitizing SQL keywords.