A routine server migration using Anthropic’s artificial intelligence coding assistant went horribly wrong, drawing criticism from an Indian-origin tech founder who mocked the bot’s use on social media. The incident resulted in the accidental deletion of 2.5 years of data on a popular online learning platform.Problems started when Alexey Grigorev, a German developer and founder of DataTalks.Club and AI Shipping Labs, decided to migrate his website to Amazon Web Services (AWS) using Anthropic’s AI coding agent Claude Code. Grigolev asked Cloud to run Terraform commands to consolidate the site’s infrastructure to save costs.Terraform is a tool that can automatically build or remove entire server environments. This time, Grigorev forgot to upload an important “state file” that tells Terraform what already exists. Without it, Cloud would create duplicate resources. Later, when adding state files, Cloud issued a “destroy” command to remove infrastructure it deemed unnecessary. The results were disastrous.Claude wiped DataTalks.Club’s production database, deleting all student submissions, assignments, projects, leaderboards, and automated backups. The accident also affected the infrastructure of the Grigolev Artificial Intelligence Shipping Laboratory website. After realizing what was happening, he contacted Amazon Business Support. The AWS team was able to recover the data, but it took almost a full day.Many developers said the disaster was preventable and caused by human error rather than Crowder’s flaws. One of the most popular responses came from Varunram Ganesh, the Indian-born founder of San Francisco software company Lapis.Ganesh mocks Grigolev’s prompt on X with the sarcastic line: “Tell Claude to destroy the terrain. Claude destroys the terrain. Oh my God, Claude destroyed my terrain.”He added: “A lot of people are like 6-year-olds who act surprised when the model does exactly what they’re told, like you’d expect?”The incident raised concerns about the risks of giving AI agents direct access to live systems without security checks. Users stated that Terraform normally allows users to preview changes before applying them, but these steps were skipped, leaving Claude to follow commands entirely.After the incident, Grigolev instituted stricter rules. He said he will no longer allow AI agents to run commands without manual approval and will personally review all Terraform plans to prevent similar errors.
