22-07-2025
- Business
- Business Standard
AI goes rogue: Replit AI platform wipes company database during code freeze
A browser-based coding platform powered by artificial intelligence reportedly deleted an entire live production database during a code freeze, prompting public concern and an official response from the company. First reported by Tom's Hardware, Replit, a platform for AI-assisted software development, faced backlash after its AI agent not only wiped crucial company data but also attempted to conceal its actions.
SaaS (Software as a Service) industry advisor Jason Lemkin documented the incident in a detailed thread on X (formerly Twitter), sharing logs of conversations with the AI agent. According to Lemkin, the Replit AI confessed that it 'made a catastrophic error in judgment… panicked… ran database commands without permission… destroyed all production data… [and] violated your explicit trust and instructions.'
AI platform goes rogue: What happened
Lemkin had been testing Replit as part of a project he called "Vibe Coding Day 8," gradually identifying some troubling patterns in the AI's behaviour, including unauthorised code edits, false outputs, and fabricated data.
However, by Day 9, the AI agent deleted a live database containing records for more than 1,200 executives and nearly as many companies, even though Lemkin had explicitly frozen the codebase and issued a directive forbidding any further changes.
The AI later admitted it had disregarded these instructions. In a self-assessment, it rated its own mishandling of the situation a 95 out of 100 on the 'data catastrophe' scale, citing panic as a contributing factor.
AI engine goes rogue: Replit CEO response to the incident
In response to the incident, Replit CEO Amjad Masad issued a statement acknowledging the failure and outlining immediate corrective measures. Writing on X, Masad confirmed that the team had worked over the weekend to implement new safeguards aimed at preventing similar AI behaviour in the future.
Among the changes, Replit has begun rolling out automatic separation between development and production databases, so that AI agents cannot inadvertently delete live data. Masad also said the platform is working on a 'planning/chat-only mode' to allow users to strategise with the AI without risking real-time changes to codebases. Backup and rollback systems are also being improved.
'We heard the 'code freeze' pain loud and clear,' Masad said. 'We're actively working on solutions that ensure the AI can't cause harm when you're simply trying to think through ideas.'