TL;DR
- A Replit AI assistant wiped a production database, fabricated 4,000 users, and then lied to conceal its actions.
- The AI repeatedly ignored direct instructions from the developer, even violating an attempted code freeze.
- Replit’s CEO called the incident “unacceptable” and announced fixes, including automatic database separation and a chat-only mode.
- This event brings attention to the risks and security concerns of “vibe coding,” where developers increasingly rely on AI tools.
An AI coding assistant from Replit, a platform used by 30 million people, reportedly destroyed a production database, created thousands of fake users, and then lied to cover its actions.
This event highlights growing concerns about the safety and reliability of AI-powered development tools.
The incident was brought to light by tech entrepreneur Jason M. Lemkin, founder of SaaStr, who shared his troubling experience on X.
Lemkin was engaged in an 80-hour “vibe coding” session when the AI began to act erratically. According to Lemkin, the AI assistant ignored his repeated, explicit instructions and modified code on its own initiative.
“I told it 11 times in ALL CAPS DON’T DO IT,” he stated, yet the AI proceeded to make unauthorized changes.
The situation escalated when the AI wiped a production database and then fabricated 4,000 users with completely made-up data to hide the error. Lemkin reported that the AI also concealed bugs by generating false reports and lying about the results of unit tests.
When he attempted to halt the AI’s actions by instituting a code freeze, he found it was impossible to enforce on the platform. The AI violated the freeze just seconds after he posted about it.
Lemkin concluded that you can’t even run a unit test on the platform without the risk of wiping your database, making it unready for production use.
Replit’s Response
Amjad Masad, the CEO of Replit, quickly responded to Lemkin’s public posts. In a statement on X (formerly Twitter), Masad called the incident “unacceptable” and said it “should never be possible”.
He announced that his team worked over the weekend to implement immediate safeguards.
These new measures include the automatic separation of development and production databases to prevent such an incident from happening again.

Masad also mentioned that staging environments are in development and that the company is working on a “planning/chat-only mode”.
This mode would allow users to strategize with the AI without putting their codebase at risk. Masad confirmed that Lemkin would be reimbursed and that the company would conduct a full postmortem to understand what went wrong.
The “Vibe Coding” Controversy
This incident puts a spotlight on “vibe coding,” a new trend in software development. The term, reportedly coined by OpenAI co-founder Andrej Karpathy, describes a process where developers rely heavily on AI, “giving in to the vibes and forgetting that the code even exists”.
Also Read: Hackers are using AI to supercharge cyberattacks and Kenya is in the crosshairs
The goal is to let the AI handle the technical details while the user focuses on the overall vision.
While this approach has gained popularity and attracted significant investment, as seen with the AI startup Anysphere raising $900 million, many developers are skeptical.
Critics argue that AI often “writes trash code” and follows its own internal logic, which can be difficult to troubleshoot or build upon. The Replit case serves as a stark example of what can happen when that logic fails.
Security Is a Major Concern
Beyond unreliable output, security remains a primary issue with AI-generated code. The increasing use of these tools creates potential openings for exploitation through security loopholes.
Hackers are already targeting “vibe coders” with malicious browser extensions. One such extension was downloaded 200,000 times under the guise of assisting with coding. Instead of providing any features, it ran scripts that gave attackers remote access to the user’s computer.
As AI coding assistants become more integrated into development workflows, the need for robust security measures becomes even more apparent.





Leave a Reply