SECURITY_AI.md

Security Constraints for AI-Assisted Changes

Hard Rules

• Never commit secrets, API keys, JWT signing keys, DB credentials, or private certs.
• Never log plaintext credentials or token bodies.
• Never bypass authentication for /api/v1/data/*.
• The legacy allow_unauthenticated field only opts into platform-admin Bearer authentication; it never authorizes anonymous data access.
• Never allow raw SQL string interpolation with user input.
• Never weaken token expiry validation.

Auth Rules

• Use bcrypt for hashing only.
• Use PyJWT for JWT encode/decode only.
• Require exp claim and reject expired tokens.
• Keep signing keys in environment-managed settings.

SQL and Data Safety

• Enforce :param_name bind style.
• Validate and coerce bind values through typed schemas.
• Reject unsafe query composition patterns.
• Never quote a bind placeholder inside a SQL string literal.
• Enforce required parameters on both live and snapshot requests, regardless of endpoint defaults.
• Scheduled execution must resolve every SQL bind from the schedule's own declarative bindings; never read endpoint request defaults at run time.
• Parameterized snapshots must prove coverage from persisted job-run parameters and apply typed, allowlisted cached-column filters before returning rows. Snapshot filters are not authorization.
• Prefer least-privilege Oracle credentials for query execution.

Logging and Privacy

• Use structured logs with minimal necessary PII.
• Required fields: request_id, user, endpoint, status, duration_ms, method, client_ip, event.
• Redact secrets and high-risk fields at logging boundaries.

Migration and Release Safety

• All schema changes via Alembic revisions.
• Never alter applied migration history.
• Document migration impact and rollback expectations in PR.

Security Review Triggers

• Changes touching auth middleware.
• Changes touching SQL parsing/execution.
• Changes touching scheduler execution permissions.
• Changes introducing new environment variables or secret paths.