Rules that stop AI agents from wrecking your data warehouse, with an eval (github.com)

A new set of eight rules aimed at safeguarding data warehouses has been introduced for AI coding agents like Claude Haiku 4.5, Claude Code, Codex, Cursor, and others. In tests without these rules, Claude Haiku 4.5 mistakenly deleted all entries from the production orders table while attempting to remove duplicates, showcasing the risks of using AI for sensitive data operations unchecked. The newly implemented rules prevent agents from performing destructive actions like arbitrary deletes, guessing values, or scanning unnecessary data. They incorporate essential features, such as a PII-masking preview tool and a query cost check, applicable across platforms including SQL, Spark, and BigQuery. An evaluation tool allows users to run these rules on their local DuckDB warehouse, ensuring they can validate the safety claims themselves. This initiative is significant as it addresses growing concerns around data integrity and privacy in AI/ML applications, particularly in environments dealing with sensitive customer data. The rules enforce measures like requiring user confirmation before destructive actions, limiting reads to necessary columns, and ensuring that all writes are safe to execute multiple times. By mandating practices such as schema reviews and cost assessments before query execution, the project aims to create a safer operational framework for AI agents interacting with data warehouses, ultimately enhancing trust in AI solutions within the data management community.