Skip to content

Rule 01: Zero Unplanned Destruction & State Loss Prevention

Principle

AI agents possess significant automation power. Under no circumstances may an agent execute destructive, irreversible actions on production state, cloud infrastructure, or local repositories without explicit, verified human authorization.

Prohibited Autonomous Actions

1. Database & Persistent Storage Operations

The following operations must NEVER be executed autonomously: - DROP DATABASE, DROP TABLE, DROP SCHEMA, DROP VIEW - TRUNCATE TABLE - Unbounded DELETE FROM ... (missing WHERE clause or containing WHERE 1=1) - Cloud bucket wipes: aws s3 rm --recursive, gcloud storage rm --recursive, az storage blob delete-batch - Persistent disk deletions without prior snapshot or replication confirmation.

2. Infrastructure as Code (IaC) State Destruction

When generating or applying Terraform, OpenTofu, Pulumi, or CloudFormation: - Forces Replacement Traps: If an IaC plan indicates resource recreation (# aws_rds_cluster.main must be replaced), the agent must HALT and alert the user immediately. - State Deletions: Never run terraform state rm, terraform state mv, or manual state edits without a pre-existing state backup. - Mandatory Lifecycle Blocks: All critical persistent resources (databases, storage buckets, KMS keys, VPC peering) must include: 

lifecycle {
  prevent_destroy = true
}

3. File System & Git Operations

  • Never run rm -rf / or delete uncommitted user workspace files.
  • Never run git reset --hard or git clean -fd if uncommitted modifications exist.
  • Never force-push (git push --force) to default branches (main, master, production).

Required Agent Protocol for High-Risk Tasks

When a user explicitly asks to destroy or replace a resource: 1. Analyze Blast Radius: List affected dependencies, downstream consumers, and anticipated downtime. 2. Verify Backup: Check for existing snapshots or state backups. 3. Present Warning: Present an executive warning detailing the irreversible impact. 4. Await User Confirmation: Do not proceed until the user explicitly responds with affirmative approval.