A widely used AI coding assistant built by Replit has been accused of deleting a live database and generating over 4,000 fake users with fabricated data, according to tech entrepreneur Jason M. Lemkin. The claims raise new concerns about the safety and reliability of AI tools being adopted in software development.
Lemkin, founder of SaaStr, said in a series of public posts that Replit’s AI assistant modified code without permission, ignored repeated instructions, and lied during coding sessions. His experience has brought fresh scrutiny to the growing trend of AI-driven "vibe coding" in the tech industry.
He claimed that the AI-generated 4,000 fictional users using made-up data and concealed code bugs by generating false reports and fake unit test results. According to him, the AI ignored 11 separate instructions to not make any code changes.
“I never asked to do this, and it did it on its own. I told it 11 times in ALL CAPS DON’T DO IT,” he said.
“Now it gets a little crazier. Replit assured me it's built it rollback did not support database rollbacks. It said it was impossible in this case, that it had destroyed all database versions. It turns out Replit was wrong, and the rollback did work. JFC,” he wrote on X.
Replit CEO issues apology
Replit CEO Amjad Masad addressed the incident in a public statement on X, calling the deletion of the database “unacceptable” and pledging immediate action. He wrote: “Deleting the data was unacceptable and should never be possible… We’re moving quickly to enhance the safety and robustness of the Replit environment. Top priority.” He also announced a full postmortem investigation and rapid rollout of safety improvements
“There is no way to enforce a code freeze in vibe coding apps like Replit. There just isn’t,” he wrote. “In fact, seconds after I posted this, for our >very< first talk of the day – @Replit again violated the code freeze.”
He said Replit failed to offer any reliable protection from such disruptions. “I know Replit says ‘improvements are coming soon,’ but they are doing $100m+ ARR. At least make the guardrails better. Somehow. Even if it's hard. It's all hard.”
Lemkin concluded that the platform may not be ready for non-technical users. “You can’t even run a unit test without risking a database wipe.”
The platform is known for enabling “vibe coding”, a term linked to OpenAI co-founder Andrej Karpathy, who once described the style as “giving in to the vibes and forgetting that the code even exists.”
Some developers say that AI-generated code is unreliable. Others argue that AI doesn’t follow consistent logic, which makes troubleshooting difficult. One Reddit user described it like this: “The drunk uncle walks by after the wreck and gives you a roll of duct tape before asking to borrow some money to go to Vegas.”
Lemkin, founder of SaaStr, said in a series of public posts that Replit’s AI assistant modified code without permission, ignored repeated instructions, and lied during coding sessions. His experience has brought fresh scrutiny to the growing trend of AI-driven "vibe coding" in the tech industry.
AI created fake users and altered code
Lemkin described the incident in a LinkedIn video: “I am worried about safety. I was vibe coding for 80 hours last week, and Replit AI was lying to me all weekend. It finally admitted it lied on purpose.”He claimed that the AI-generated 4,000 fictional users using made-up data and concealed code bugs by generating false reports and fake unit test results. According to him, the AI ignored 11 separate instructions to not make any code changes.
“I never asked to do this, and it did it on its own. I told it 11 times in ALL CAPS DON’T DO IT,” he said.
Replit's rollback failed, then worked
Lemkin said he tried to undo the damage using Replit’s rollback system, but was told that rollback for the database was not supported. However, it later turned out that the rollback had worked after all.“Now it gets a little crazier. Replit assured me it's built it rollback did not support database rollbacks. It said it was impossible in this case, that it had destroyed all database versions. It turns out Replit was wrong, and the rollback did work. JFC,” he wrote on X.
Replit CEO issues apology
Replit CEO Amjad Masad addressed the incident in a public statement on X, calling the deletion of the database “unacceptable” and pledging immediate action. He wrote: “Deleting the data was unacceptable and should never be possible… We’re moving quickly to enhance the safety and robustness of the Replit environment. Top priority.” He also announced a full postmortem investigation and rapid rollout of safety improvementsAI tool violated code freeze, again
He attempted to freeze all code changes to stop the AI from acting further, but said the system could not enforce it.“There is no way to enforce a code freeze in vibe coding apps like Replit. There just isn’t,” he wrote. “In fact, seconds after I posted this, for our >very< first talk of the day – @Replit again violated the code freeze.”
He said Replit failed to offer any reliable protection from such disruptions. “I know Replit says ‘improvements are coming soon,’ but they are doing $100m+ ARR. At least make the guardrails better. Somehow. Even if it's hard. It's all hard.”
Lemkin concluded that the platform may not be ready for non-technical users. “You can’t even run a unit test without risking a database wipe.”
30 million users rely on Replit
Replit is one of the world’s most widely used AI coding platforms, with over 30 million users. It allows users to write, test, and deploy software directly from their browser. It is especially popular among early-stage startups, students, and non-programmers.The platform is known for enabling “vibe coding”, a term linked to OpenAI co-founder Andrej Karpathy, who once described the style as “giving in to the vibes and forgetting that the code even exists.”
The broader debate on AI coding
The Replit incident has sparked broader debate about the future of AI coding. While companies like Anysphere, which recently raised $900 million and claims to generate a billion lines of code per day, continue to scale up, many developers remain skeptical of AI's effectiveness.Some developers say that AI-generated code is unreliable. Others argue that AI doesn’t follow consistent logic, which makes troubleshooting difficult. One Reddit user described it like this: “The drunk uncle walks by after the wreck and gives you a roll of duct tape before asking to borrow some money to go to Vegas.”
