How do I use QueryShield with MotherDuck / DuckDB?
DuckDB and MotherDuck (DuckDB-in-the-cloud) are an increasingly popular substrate for AI analytics agents because of their fast columnar engine and SQL ergonomics. QueryShield uses sqlglot's DuckDB dialect parser.
import duckdb
from queryshield import Clientqs = Client(api_key=KEY, policy="duck-analytics", dialect="duckdb") con = duckdb.connect("md:my_database?motherduck_token=...")
decision = qs.validate(sql=llm_output, context=ctx.dict()) if decision.allowed: df = con.execute(decision.sql).df() ```
DuckDB-specific considerations:
1. COPY ... TO 'file.parquet', EXPORT DATABASE statements can write to the host filesystem — denylist for any LLM-driven agent.
2. INSTALL, LOAD of extensions — denylist; an attacker steering the agent to load httpfs and then COPY to a remote S3 bucket is a realistic exfil path.
3. ATTACH 'remote.db' — denylist; can pull in arbitrary databases.
4. MotherDuck shares — policy can restrict the agent to a single database namespace (USE my_db enforced; cross-database queries denied).
5. DuckDB has no traditional RLS — the AST layer is your primary control, not a defense-in-depth complement.
The lack of native RLS makes QueryShield's policy DSL especially load-bearing on DuckDB; required predicates are the only enforcement layer for multi-tenant data.