database-security

The database enforces what application bugs forget: least privilege, read-only by default for agents, no destructive queries without explicit authorization.

Activate when

• Creating roles, connection strings, or MCP/database tool access.
• Letting an agent query production or staging data.
• Reviewing migrations touching permissions or sensitive columns.

Do NOT activate for

• Schema/query performance (see postgres-operations).

Procedure

1. Roles: app role gets only the statements it needs on only its schema; separate migration role; no superuser for apps or agents.
2. Agent/MCP access: read-only connection, explicit table/column allowlist, row cap + statement timeout, query validation before execution; block DROP/DELETE/UPDATE/TRUNCATE/ALTER/GRANT unless the task explicitly authorizes a scoped write with owner sign-off.
3. Secrets: credentials from the secret manager, never code or logs; rotate on exposure.
4. Sensitive data: encrypt in transit and at rest, minimize logged fields, mask PII in traces and error messages.
5. Verify: connect as the app/agent role and confirm a forbidden statement fails; confirm destructive patterns are rejected by the guardrail, not just by policy text.

References

• ../postgres-operations/ for operational procedures.
• ../../security/secrets-supply-chain/ for secret handling.