Base.metadata.drop_all(engine) drops every table in the metadata that exists in the database the engine points at. It is a normal teardown step in tests, and it destroys data when the engine resolves to production, usually because the test process loaded the production .env or DATABASE_URL. If that happened, stop the test runner and the services, and restore to a moment just before the run: create_all() recreates the tables, not the rows.
Recover now
- Stop the agent or the process. End the agent session, cancel any background task, and stop the test run, deploy job or service that ran the command, so nothing runs again against the database you are about to restore.
- Do not run more migrations, pushes or resets, and do not let the agent "repair" the schema. Recreating the tables brings back empty tables, not the rows. Write down when the destructive command ran (from the agent transcript, the shell history or the database log): every restore below needs that moment.
- Restore to a point just before the command, into a separate instance. Self-managed PostgreSQL: restore the base backup and replay the archived WAL with
recovery_target_timeset just before it; apg_dumpfile restores only to when it was taken. Amazon RDS / Aurora: "Restore to point in time" creates a new DB instance (RDS) or cluster (Aurora) at any moment within the backup retention period, without modifying the source. Neon: instant restore takes a root branch back to a moment within its history window. Supabase: the daily backups and, where it is enabled, point-in-time recovery are under Database → Backups. Railway: volume backups, if they were created or scheduled, are restored from the service. MySQL: restore the last full backup, then replay the binary log withmysqlbinlogup to just before the destructive event. - Check the restored copy, then move it back. Compare row counts and the latest rows with what you expect, then point the application at the restored instance or copy the tables back (
pg_dump -t tablefrom the restored instance). Rows written between the restore point and now must be re-applied by hand. - No backup and no point-in-time recovery? Look for other copies: a staging or development database, a read replica, exports, a data warehouse, or the systems the data came from.
Why it happened
Per the SQLAlchemy docs, MetaData.drop_all() drops all tables stored in the metadata and, by default (checkfirst=True), skips the ones not present in the target database; tables=[…] limits it to a subset. It has no notion of environment: it drops on whatever engine it is given. The engine usually comes from an environment variable, so a test run with the production .env loaded, from a deployment directory or a shared shell, tears down production. A developer reported exactly that with an AI agent running the tests (openai/codex#42480).
Make the test fixture refuse production
Give tests their own variable and make the fixture check where it is about to drop tables, so a wrong environment fails before the teardown:
import os
import pytest
from sqlalchemy import create_engine
@pytest.fixture
def engine():
url = os.environ["TEST_DATABASE_URL"] # never DATABASE_URL
engine = create_engine(url)
if not (engine.url.database or "").endswith("_test"):
pytest.exit(f"refusing to drop tables in {engine.url.database}")
Base.metadata.create_all(engine)
yield engine
Base.metadata.drop_all(engine)This protects the fixtures you write. It does not protect a test someone else wrote, or one an agent writes, which is why the connection itself should not reach production.
How to prevent it
drop_all() sends one DROP TABLE per table over the engine's connection. When the only path to production is the proxy, those statements are blocked before they reach the database, so a test run that resolves to production fails instead of emptying it.
Vericto checks each SQL statement on its AST before it reaches the database. These are the rules that report the statements of this incident, with the real CLI output on each rule's example:
VERICTO-010: DROP TABLE or DROP DATABASE
The rule's example · Critical
-- migrations/0042_drop_users.sql
DROP TABLE users;
Real CLI output
$ npx @vericto/vericto-cli@2.3.0 check --offline migrations/0042_drop_users.sql
note: offline mode: 42 standard rules at default severities, engine 5.1.0. Custom rules, overrides, sensitive columns, allowlists and the audit trail apply online only.
! migrations/0042_drop_users.sql FLAGGED VERICTO-010 critical
DropStmt
fix: Use a versioned migration tool (Flyway, Prisma Migrate, Alembic, Rails migrations)
0 blocked · 0 allowed · 1 flagged · 0 monitored (ruleset standard@engine-5.1.0 · mode: offline (engine 5.1.0, 42 standard rules, no workspace policy))
In a migration file the CLI marks it FLAGGED and exits with 0: a schema change in a migration is usually intended, so it is raised for review instead of blocked. With --fail-on flag (or fail_on = "flag" in .vericto.toml) it exits with 1 and the job fails. Outside a migration (--sql, stdin or another file) it blocks, as the proxy and MCP check_sql do.
Set up the agent so it goes through Vericto
- Connect the agent to the database through the proxy. Its connection URL points at the Vericto TCP proxy, not at the database. The proxy checks each statement, including the ones an ORM command run by the agent sends, and at runtime it blocks
DROP,TRUNCATEandALTER … DROP COLUMNbefore they reach the database. For a coding agent,vericto-proxy mcpgives itexecute_sql,describe_schemaandcheck_sql, never the database credential. - Keep the direct production credential away from the agent. Not in the
.env, the shell or the environment the agent and its test runs load. If the agent has another way to the database, the proxy does not see it. - Give it a read-only role through Vericto. The role behind
vericto-proxy mcpis read-only (the SQL to create it is in the MCP docs), and agent access sets each agent to read-only on the tables you choose. - Let the agent check SQL before it runs it with the MCP
check_sqltool: it gets the verdict, the rule and the suggested fix.
In CI, on the migration files
A test fixture produces no .sql file, so CI cannot see a drop_all(). Your Alembic migrations can be checked: alembic upgrade head --sql writes the SQL they would run (offline mode), and the Action checks that file in the pull request.
The vericto/check-sql GitHub Action (Apache-2.0) runs the same offline check on the .sql files a PR changes. Save it as .github/workflows/vericto.yml:
on: pull_request
jobs:
vericto:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v7
with: { fetch-depth: 0 }
- uses: vericto/check-sql@v1
with: { offline: true, upload-sarif: false }
Or try the CLI on your machine, with no account (AGPL-3.0-only):
npx @vericto/vericto-cli@2.3.0 check --offline migrations/
Scope: Vericto protects the connections that go through Vericto. A shell on the database host, or a credential that connects directly, bypasses it.
Frequently asked questions
Sources
The public reports this page cites, checked on 2026-11-14, and the official docs behind each fact about the tools:
- openai/codex#42480 (September 3, 2026): a developer reports a test whose setup called Base.metadata.drop_all(engine) while the production .env was loaded
- SQLAlchemy 2.0 docs: MetaData.drop_all()
- Alembic docs: generating SQL scripts (offline mode)
- PostgreSQL docs: continuous archiving and point-in-time recovery
- Amazon RDS docs: restoring a DB instance to a specified time
- Amazon Aurora docs: restoring a DB cluster to a specified time
- MySQL 8.4 manual: point-in-time recovery using the binary log