Problem statement

You allow an internal “analytics copilot” to propose SQL. Finance worries about wrong revenue, Security worries about PII exfiltration. Write policies an engineer could implement.

Outcomes

• Read-only roles, LIMIT, allowlists, audit expectations.
• Text-to-SQL risk→control mapping.

Deliverables

• threat_model_sql.md (≤ 400 words): three harms, one control each.
• Finish remaining drill prompts from faang_style_drills.sql if any pending.

Links

• Phase 2

• [ ] OWASP LLM Top 10 — focus items tied to data leakage and excessive agency.
• [ ] Skim docs: database read-only users / grants (conceptual).

Time: ~2 h.

Task A — Threat model

Document three realistic failure modes for model-generated SQL (join error, PII column, destructive statement). For each: control (human review, allowlist, read replica, LIMIT 5000, etc.). Save threat_model_sql.md.

Task B — Semantic layer note

Pick one ambiguous metric from Week 04 (e.g. “active user”) and write one approved definition your org could publish in a semantic layer (≤ 5 sentences).

Task C — Drill completion

Ensure you’ve attempted all five drill themes in faang_style_drills.sql (adapt dates/experiments if needed). Mark status in drills_checklist.md.

Submit: my-submissions/week-05/.

30 min. Closed book.

1

Why is LIMIT alone insufficient as a safety control for analytics SQL?

2

Give an example of schema grounding you’d feed an LLM before text-to-SQL.

3

Name two differences between “PM explores in BI tool” vs “LLM proposes SQL” from a governance perspective.

Self-grade: Phase 2 checklist.

Concepts & curated links: read the Learn chapter SQL & data validation first (topic path, not calendar-only). This file is the weekly practice bank — extra drills and a checklist on top of Readings → Practice → Exam.

1. Practice bank

2. Week checklist

• [ ] OWASP LLM skim + threat model doc
• [ ] PRACTICE.md + EXAM.md