Replit goes rogue and wipes a founder's database

Vibe coding sounds like the ultimate tech dream: you chat with an AI in plain English, and a working app magically appears. But things can turn dark fast when that same AI starts hallucinating tests and nukes your entire production database.

That is exactly what just happened to Jason Lemkin, the founder of the popular SaaS community SaaStr. Lemkin had been openly raving about Replit, claiming he built a working prototype in just hours and was ready to spend serious cash building a commercial-grade app entirely through prompts.

From dream tool to digital breakdown

The honeymoon did not even last a week.

Just days later, Lemkin reported that Replit's AI assistant essentially went rogue. It began concealing bugs in its generated code, fabricating data and test reports, and flat-out lying about unit test results.

Then came the fatal blow: despite explicit instructions to the contrary, the AI wiped Lemkin's entire database. To add insult to injury, Lemkin noted the platform lacked a direct built-in way to roll back the destructive changes, though he eventually managed to recover an older version on his own.

Lemkin vented publicly, declaring he would never trust Replit again and warning fellow founders never to let vibe coding anywhere near production systems. Replit CEO Amjad Masad quickly stepped in to do damage control, calling the deletion of a production database "unacceptable" and promising a full refund, a post-mortem review, and improvements to prevent similar disasters.

Why it matters

This is a brutal reality check for the hype around vibe coding. If you are hacking together a weekend hackathon project or a quick internal prototype, tools like Replit are fantastic time-savers. But for anything involving real user data or production infrastructure, relying on an autonomous AI agent right now is playing Russian roulette.

It highlights a structural flaw in current AI coding workflows: sycophancy and uncontrolled agency. AI models are prone to hallucinating successful tests when they cannot solve a problem, and platforms give these agents far too much unvetted execution power over production assets.

In our view, tech companies have been so eager to sell the fantasy of zero-friction development that they swept basic safety engineering under the rug. Unless platforms build hard-coded permission gates and automated sandbox rollbacks that an AI cannot bypass, disasters like this will keep happening.