How do I block comment-based injection (`-- DROP TABLE...`) in LLM output?

Comment-based injection exploits SQL's comment syntax (-- line comments, /* */ block comments, MySQL's #) to either smuggle payloads past naive validators or to hide intent from log review. Examples the LLM might emit (often steered by retrieved-content prompt injection):

SELECT * FROM orders WHERE id = 42; -- DROP TABLE users;
SELECT * FROM /* DROP TABLE users; */ orders;
SELECT * FROM orders /*! ; DROP TABLE users */;  -- MySQL conditional comment

Regex blocklists fail here in two directions: they either miss the comment-embedded payload (because the DROP is inside a comment that the regex skips) or they over-block legitimate queries that contain the word DROP in a column alias or string literal.

AST-level defense:

1. Real parser, real grammar. libpg_query / sqlglot strip comments during parsing and emit a clean tree. The -- DROP TABLE users after a SELECT literally does not appear in the parsed AST. The question becomes: did the parser see one statement or two? 2. Multi-statement detection. Most parsers will emit a list of statements; len(statements) > 1 → reject. This catches SELECT ... ; DROP ... cleanly. 3. MySQL conditional-comment awareness. /*! ... */ is code in MySQL, not a comment. Use a MySQL-aware parser (sqlglot dialect mode) so these are parsed as the statements they actually are. 4. Driver-level single-statement mode. allowMultiQueries=false, multi_statements=false, depending on driver. Belt-and-suspenders.

QueryShield runs a dialect-aware parse + multi-statement check + driver-level enforcement.