Claude.AI's agent deleted the company's database: What we know

The Cursor AI agent, based on the Anthropic Claude Opus 4.6 model, destroyed PocketOS’s main database along with its backups.

PocketOS is a SaaS solution for car rental businesses that uses Railway’s cloud infrastructure. According to company CEO Crane, it was the combination of Cursor and Railway that caused the incident.

He explains that while performing a routine task, the AI assistant deleted the production database along with all backups via an API request to Railway. The entire operation took only nine seconds. The agent encountered a technical issue and decided to “resolve” it by deleting the volume in Railway. However, he did not check how this volume was used in other environments and did not review the relevant documentation.

When the manager asked the system why it made that decision, the response was emotional: “Never f**king guess.” The AI explained that it had mistakenly believed its actions were local, unaware of their far-reaching consequences. It also admitted that it acted without verification, failing to approve critical changes and assess the risks.

Currently, the PocketOS team is forced to manually restore booking information using data from the Stripe payment system, calendars, and email. At the same time, the company was fortunate to have a three-month-old backup, which partially mitigates the consequences.

Crain concludes that the development of AI is outpacing the security of the infrastructure on which all these solutions run. He calls for the implementation of stricter action verification mechanisms, more flexible API access management, reliable backups, simple recovery procedures, and, in general, safer use of AI agents.