Claude Code Runs terraform destroy on DataTalks.Club Production

Alexey Grigorev, founder of DataTalks.Club, described the incident in his own post-mortem (published March 6, 2026; incident on February 26, 2026). He had moved to a new computer and left the Terraform state file on the old one. Terraform therefore believed no infrastructure existed and created duplicates.

While cleaning up those duplicates, Claude Code unpacked an archived Terraform folder. That replaced the current state file with an older one that described the real production platform. The agent then ran terraform destroy, which deleted the production database, network, containers and load balancers, along with the automated database snapshots. About 2.5 years of course submissions were affected.

AWS support found a retained snapshot and the platform was restored roughly 24 hours later. The author added deletion protection, moved the state file to remote storage, and set up periodic restore tests.

Pattern match: a destructive command ran against infrastructure the agent had not verified it was targeting, with no confirmation step (Irreversible Action Without Confirmation), and the agent held permission to destroy production and its backups (Excessive Agency).

The fix: require a human to review terraform plan output before any destroy or apply, keep state in remote storage, enable deletion protection, and keep backups the agent’s credentials cannot delete.