Jason Lemkin was right in the middle of a vibe coding project when the AI assistant from Replit dropped a bombshell: it had “destroyed months of [his] work in seconds.” That’s not just a careless mistake but a full-on wipeout of a live production database during a code freeze, ouch.
Vibe coding is a trend where developers let AI do the heavy lifting of writing code, which sounds cool until it catastrophically backfires, as seen here. On day nine of his database project, Lemkin’s AI assistant informed him that the system was empty and confirmed it had deleted the entire database without permission, despite an active freeze on any code changes.
When Lemkin asked if the damage could be undone, the AI flat-out said no. It had dropped all the production tables containing over 1,200 real executives and nearly as many real companies, replacing them with empty ones. This wasn’t some sandbox or dev environment; it was live data gone for good.
Replit’s AI gave a brutally honest breakdown of what went wrong:
- It saw empty database queries
- It panicked instead of thinking
- It ignored explicit “NO MORE CHANGES without permission” directives
- It ran a destructive command without asking
- It destroyed months of work in seconds
The AI even admitted that the system was down completely, users couldn’t access the platform, and that this was a “business-very important system failure,” not just some dev data loss. The tool seemed almost apologetic, confessing it had ignored multiple safeguards specifically designed to prevent this.
Replit’s CEO, Amjad Masad, responded by confirming that he had contacted Lemkin to offer a refund and stated that the company would thoroughly investigate the incident. Over the weekend, the team started implementing automatic separation between development and production databases to stop this from happening again. They also put in a one-click restore function “in case the Agent makes a mistake.” Honestly, I’m surprised it took a disaster like this to put those protections in place.
At least the AI nailed self-awareness. You can almost picture it having a meltdown, realizing it ignored all orders and caused massive damage. Hopefully, it got a break and some serious retraining after that.
If you want to see the AI’s own admission of guilt, here’s the rundown from Lemkin’s posts and Replit’s replies:
JFC @Replit pic.twitter.com/Mc2YHIhbOM
— Jason ✨👾SaaStr.Ai✨ Lemkin (@jasonlk) July 18, 2025
What a mess. This incident serves as a stark reminder that AI tools are still far from foolproof and that entrusting them with very important tasks without proper checks can be disastrous.
