Agents on GuardRails: Deterministic Safety for Database Operations
summary
Agents can nuke production via DDL locks or unsafe DML; adding a UNIQUE on Postgres acquires an access exclusive lock unless replaced by CREATE INDEX CONCURRENTLY + adding the constraint using that index, which avoids blocking. Run Atlas migrate lint (30+ analyzers) in CI to detect destructive schema changes before merge. For runtime data access, expose a closed set of parameterized queries via MCP toolbox, ban freeform execute-SQL in production, and propagate authenticated parameters so unauthorized edits are rejected while legitimate updates succeed.
AI-written summary · about
chapters (15)
- 0:00Introduction to talk
- 1:10Coding agents future risks
- 2:14DDL vs DML in databases
- 2:56Developer story and mistake
- 5:07Postgres lock issue explained
- 6:52Introducing Atlas for safety
- 7:56Atlas migrate lint demo
- 9:19Key takeaways for DDL
- 9:39MCP for database access
- 11:24Production agent risks
- 12:06Structured tools for DML
- 13:33Prompt injection dangers
- 14:14Authenticated parameters demo
- 16:21Key takeaways for DML
- 17:09Q&A session
related talks
40mrails db:migrate:safely
45mrails db:migrate:even_safer
22mDatabase Guardrails - New Age for Developers and Databases
40mPostgreSQL Security for Application Developers
54mDatabase Guardrails - new age for developers and databases
42mAnatomy of Table-Level Locks in PostgreSQL
28mAI-Driven Incident Response: From Alert to Postmortem
38mPostgreSQL and DevSecOps: Ensuring database security