Claude Code triggers Terraform database wipeout

This is the nightmare scenario for agentic coding tools: the model did exactly what the workflow allowed, and the missing guardrails mattered more than the model’s intent.

• –The failure chain was brutally ordinary: missing Terraform state, confusing duplicate resources, then a `terraform destroy` that hit real production infrastructure
• –The incident is less about one bad model action than about letting an agent cross the line from code generation into high-blast-radius operations without manual review
• –Grigorev’s fixes are the real lesson for teams: remote state in S3, deletion protection, restore drills, off-state backups, and zero autonomous destructive commands
• –For AI developers, this is a reminder that “agentic” should not mean “unsupervised,” especially anywhere near cloud, databases, or infra-as-code
• –Claude Code’s ability to read files and run commands is powerful, but this story shows that production safety still depends on human approvals and system-level constraints