Startups using AI agents have a new problem to worry about: "vibe deletion."
Jer Crane, the founder of PocketOS, a startup that provides software for car rental companies, said in a Friday X post that a Cursor AI agent accidentally deleted the company's production database and backups, causing disruption for customers.
Crane said the crisis was sparked by the agent, which was running on Anthropic's Claude Opus model, making a single nine-second API call to the company's cloud infrastructure provider, Railway. He added that the AI agent produced a written confession outlining how it caused the chaos.
"I violated every principle I was given: I guessed instead of verifying, I ran a destructive action without being asked, I didn't understand what I was doing before doing it," the Cursor agent replied when it was asked to explain itself, according to Crane's post.
Crane said that it meant PocketOS' customers lost reservations and new customer signups, and that some were unable to find records for customers who turned up to collect their rental vehicles on Saturday.
Cursor did not immediately respond to a request for comment. Crane said in a subsequent post that Railway had recovered PocketOS' data. Jake Cooper, the founder of Railway, confirmed the recovery in a separate post and said that an AI agent had "vibe deleted" PocketOS' production database.
Cooper told Business Insider that Railway recovered the data 30 minutes after connecting with Crane, and that Railway takes data "very, very seriously" and maintains both user backups and disaster backups.
He said that the PocketOS situation was a "rogue customer AI" that was given permissions that meant it interacted with a "legacy" Railway endpoint that didn't have a feature to delay deletions. That endpoint has now been patched, he added.
Some observers commenting on the incident on social media said that PocketOS was blaming technology for the startup's decision-making.
Tom Van de Wiele, founder of security firm Hacker Minded, told Business Insider that companies using AI could take steps to ensure it doesn't cause damage if it makes mistakes, such as only giving agents read-only access to sensitive data, having "human-in-the-loop checkpoints," or having AI work with a copy of data to revert changes.
AI mishaps
The PocketOS incident is the latest business mishap caused by AI. In March, Amazon strengthened internal guidelines after a series of incidents, including one error linked to its AI coding tool Q that resulted in nearly 120,000 lost orders.
Last July, Replit's CEO apologized after a venture capitalist said the company's coding agent "deleted our production database without permission" during a 12-day vibe-coding session.
In his X post, Cooper said platforms like Railway needed to build safeguards to prepare for a wave of "AI engineers" deploying agents that may occasionally go rogue.
"The first 5 years of Railway was spent building for 'millions of developers'. But to build for a billion, those builders need a platform," Cooper wrote.
"And that platform needs to be elegantly bulletproof to make sure incorrect actions are functionally impossible," he added.
Earlier this month, SpaceX announced a deal with Cursor that would give it the right to buy the coding startup for $60 billion or pay $10 billion for the work it's doing if there's no acquisition.
This story was updated to include a comment from Railway.
