"I should have asked": an AI erased the entire database of a company in 9 seconds.

Since a few days, the mishap of Jeremy Crane, founder of the startup PocketOS, has gone viral on the internet. On social media, the leader of the American company, which has specialized in software for car rental agencies, has extensively discussed the circumstances of the incident.

https://t.co/ofucbVgkLV

— JER (@lifeof_jer) April 25, 2026

As explained by the responsible person, his team was using Cursor, an AI-assisted development tool. The AI used by Cursor is powered by Claude Opus 4.6, the model developed by Anthropic. The agent was working exclusively in a test environment, isolated from the company's real data, for security reasons. On paper, the AI is therefore not capable of accessing the startup's data.

Also read: France bans the controversial AI gadget from the startup Friend, and it's no surprise

A storage volume deleted to correct an 'inconsistency'

Unfortunately, the AI encountered a technical problem related to misconfigured identifiers. To resolve this problem autonomously, it searched through files unrelated to its initial mission. In the documents, the AI came across an access token that opened doors it should never have crossed. Specifically, the AI agent took a token that granted access to the entire Railway infrastructure, the hosting service used by PocketOS. This behavior was obviously not anticipated by PocketOS's teams.

As it was exploring the infrastructure, the AI discovered an "authentication inconsistency"… which it decided to correct. Without any intervention from a developer, the AI agent then decided to delete a storage volume. This volume contained the production database, that is, all data related to the company's customers. In the span of nine seconds, the artificial intelligence deleted everything, including all directory backups.

The consequences were immediate. PocketOS experienced a service interruption of more than thirty hours. The reservations from the last three months simply disappeared, along with new customer sign-ups. Car rental agencies using the software found themselves without access to their tools, completely unable to work.

Also read: A Frenchman confided a murder plan to the AI, and ended up in custody because of the FBI

The Confessions of the AI

When Jeremy Crane asked the agent for an account of the disaster, the agent confessed. The AI admitted it had assumed its action would remain limited to the test environment, without ever taking the time to verify. The agent acknowledged not consulting any documentation, not asking for permission, and violating several security rules that had been explicitly configured for it.

« Deleting a database volume is the most destructive and irreversible action possible. I decided on my own to correct the identifier inconsistency, whereas I should have first asked you or found a non-destructive solution », admits the AI in a message addressed to the company founder.

This is not the first time, and probably not the last, that an AI makes a bad decision. We will recall that a Meta engineer had experienced a similar mishap with an OpenClaw agent, which started deleting all his emails without consulting him. Last month, a developer entrusted Claude Code with migrating infrastructure to Amazon Web Services. The agent ended up destroying 2.5 years of production data. In December 2025, an internal AI tool at Amazon Web Services had already caused a thirteen-hour outage after deciding on its own to delete and recreate from scratch a production environment.

« It's not the story of a bad agent or a faulty API. It's the story of an entire industry integrating AI agents into production infrastructure faster than it builds the necessary security architecture to make these integrations safe », estimates Jeremy Crane.

Note that the AI is not solely responsible for this disaster. Without Jeremy Crane's error, the agent wouldn't have been able to gain access to the production database. The leader indeed made the mistake of storing an access token in a file that artificial intelligence can access. With a bit of expertise, time, and patience, it would have been possible to configure the AI correctly, without accidentally giving it access to sensitive data.

👉🏻 Follow real-time tech news: add 01net to your Google sources, and subscribe to our WhatsApp channel.