SQLAlchemy 2.0

Apply the Python 3 baseline first. These rules cover SQLAlchemy 2.0 query construction, textual SQL, sessions, transactions, tenant isolation, connection security, logging, and migrations.

Values and textual SQL

• Build queries with SQLAlchemy expressions and bind every attacker-controlled value as a parameter.
• When text() is necessary, keep the SQL string constant and declare values with bindparams() or the execution parameter mapping; placeholders cannot bind identifiers or SQL syntax.
• Never interpolate or concatenate untrusted values into SQL, including through f-strings, %, format, or template engines.
• Treat literal_column, text, from_statement, raw driver SQL, execution options, and bindparam(..., literal_execute=True) as reviewed escape hatches. Never pass untrusted values to literal_execute parameters, which render inline at execution time, and never execute SQL compiled with literal_binds=True from untrusted values.
• Do not use escaping or manual quoting as a substitute for bound parameters.
• Escape LIKE metacharacters in untrusted search terms by calling contains(), startswith(), endswith(), or their case-insensitive icontains(), istartswith(), and iendswith() forms with autoescape=True; do not build like(f"%{term}%") or ilike(...) patterns from raw input. Never use like() or ilike() for identity lookups such as email or username matching; compare func.lower(column) == value.lower() or use a case-insensitive collation instead.

Identifiers and dynamic queries

• Map client-selected sort, filter, column,