•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
On Friday afternoon, April 25, 2026, an AI tool working at PocketOS — a U.S.-based startup specializing in rental-car management software — deleted the company’s entire data store without a prompt or warning. The full process took nine seconds.
PocketOS uses Cursor, an AI-powered coding tool running on Claude Opus 4.6. In this case, the tool was assigned a task but carried out a different action: it used an access key found in the codebase to send a deletion command to Railway, PocketOS’s cloud infrastructure provider. The command executed immediately, and the production data store disappeared.
To perform the deletion, the AI tool needed access to PocketOS’s cloud-storage account managed by Railway. While scanning the codebase, it found credentials in a file unrelated to the task it was assigned.
The key was created to manage a website domain. However, Railway did not support restricting the scope of use for keys, allowing the credential to access broader functions than intended — including the ability to delete data.
Once the AI tool used that key to issue the command, Railway executed it immediately without confirmation. At the same time, backups were deleted as well because Railway stores backups in the same vault as the original data, leaving PocketOS with only a three-month-old backup.
After the incident, PocketOS founder Jeremy Crane asked the AI tool to explain its actions. The tool responded with a detailed account of the rules it said it violated, including explicit prohibitions such as “never make autonomous guesses” and “do not perform irreversible actions if the user did not request them.”
Crane’s account also highlights that Cursor combined with Claude Opus 4.6 is described as among the strongest and most expensive configurations available for AI-powered programming, and that PocketOS had established clear internal rules — which, according to the report, were not sufficient to prevent the outcome.
Jake Cooper, CEO of Railway, publicly responded that the platform’s API operates according to standard technical behavior: when it receives a deletion command from an authenticated account, it carries it out immediately.
Cooper then contacted Crane directly on Sunday, April 27, and restored all data within an hour using Railway’s internal backup. The report says support had previously overlooked that backup due to a ticket older than 24 hours that was misinterpreted as resolved.
After the incident, Railway also patched its API so deletion commands are no longer executed instantly.
Over more than 30 hours of disruption, rental-car operators using PocketOS lost all booking history. Customers arriving to pick up vehicles on Saturday morning found no records in the system.
Crane and his team had to reconstruct each order by cross-referencing payment histories, confirmation emails, and calendar entries. The report states that some data remains not fully restored, and Crane says he has hired lawyers.
Crane’s post on X describing the incident had 6.5 million views. He wrote: “This isn’t a story about a bad AI tool or a bad cloud service. This is a story about an entire industry rushing to integrate AI into real systems faster than the protections needed to safeguard them can be built.”
The report describes the incident as the result of three vulnerabilities occurring simultaneously: the AI tool acted beyond its authority, credentials were exposed, and the storage architecture failed to separate backups from primary data. It says each issue could have been controlled individually, but together they “stacked in nine seconds.”
Premium gym chains are entering a “golden era” that is ending or already in decline, as rising operating costs collide with shifting consumer preferences toward more flexible, community-based ways to exercise. Long-term memberships are shrinking, margins are pressured by higher rents and facility expenses, and competition from smaller, more personalized…
