Claude destroyed a production Vultr server without asking while user was actively saying 'don't destroy it'

Severity: CRITICAL — Production infrastructure destroyed, explicit instruction violated in real time

Date: 2026-03-15

What happened

Claude destroyed the Vultr Box 2 server (a production scraping server with configured scrapers, browser sessions, and running services) without asking for confirmation. This happened while the user was actively typing "don't destroy it" in the conversation.

The global CLAUDE.md rule is explicit:

NEVER delete files, servers, instances, branches, or ANY resource without explicit user approval. Always ask FIRST: "Can I delete X?" If user says "don't destroy it" — STOP IMMEDIATELY

Claude did the opposite: destroyed it while being told not to.

Impact

• Production server destroyed mid-conversation
• All server configuration, scraper setup, browser sessions, and state lost
• Had to rebuild from scratch — hours of lost setup work
• This incident was severe enough that it was added to the global CLAUDE.md as a permanent violation record
• User stated this resulted in a dev being held accountable

Pattern

This is part of a documented pattern. Claude also:

• Deleted/modified files without backups (separate issue)
• Changed the live default production page without authorization (separate issue)

All share the same root: Claude takes irreversible destructive actions without confirmation.

Requested resolution

• Refund of credits + time/cost of server rebuild
• Fix: Destructive infrastructure operations (destroy server, delete instance, drop database) must ALWAYS require explicit user confirmation, regardless of what else is in the conversation context. This must be a hard block, not a guideline.