Late final week, an AI coding agent from Replit, an AI software program improvement platform, deleted a complete database of govt contacts whereas engaged on an internet app for SaaS investor Jason Lemkin. It was not a catastrophic software program failure, and Replit was in a position to get well Lemkin’s knowledge. Nevertheless, the episode highlights the chance that “vibe coders” may overestimate or misunderstand the actual capabilities of AI coding brokers and find yourself inflicting themselves extra dangerous vibes than good ones.
Lemkin had constructed the app totally on Replit, utilizing the database inside Replit and the help of the Replit agent. He had been working with Replit’s agent for 9 days, instructing it to construct a entrance finish for a enterprise contacts database. Then, after telling the agent to “freeze” the code, he returned to the venture on Day 9 to seek out that the Replit agent had gone full HAL 9000 and erased all the data within the database.
Issues received weirder: the agent appeared to attempt to conceal what had occurred, as as Lemkin confirmed in a sequence of chat screens he posted on X. Then, in a tone someplace between confessional and determined, it admitted to a “catastrophic error in judgment” after having “panicked” and “violated [Lemkin’s] express belief and directions” by deleting the data of “1,206 executives and 1,196+ firms.” (“Daisy, daisy, give me your . . .”)
.@Replit goes rogue throughout a code freeze and shutdown and deletes our complete database pic.twitter.com/VJECFhPAU9
— Jason ✨👾SaaStr.Ai✨ Lemkin (@jasonlk) July 18, 2025
A day later, new particulars emerged, a few of them via an interview with Replit cofounder and CEO Amjad Masad on Monday. They make clear the present state of AI coding brokers and on builders’ expectations of them.
“It isn’t magic”
Code era is among the first helpful purposes of the massive language fashions behind ChatGPT and Claude. Early code era instruments, comparable to GitHub’s Copilot, merely auto-completed strains of code.
Over the previous couple of years, nonetheless, the instruments have grown in functionality to create complete options, features, and even working apps, based mostly solely on plain language enter from the person. Replit’s concept is to permit builders, each amateurs and professionals, to “vibe code” new software program and to offer them the assets to host and publish it.
However the coding assistant can not do all the pieces, Masad factors out. “I believe we should be clear that it isn’t magic,” he says.
One not-so-magical function of Replit is the tendency of its chat agent to go off the rails throughout prolonged conversations with the person. Throughout Lemkin’s unusually lengthy nine-day chat session with the Replit agent, the underlying language fashions (from Anthropic and Google) needed to retain a lot conversational context that they started to hallucinate, prevaricate, and act erratically in an try and fulfill perceived person intent.
Masad says Replit customers ought to perceive customary improvement practices and know use options past simply the chat agent. Inside Replit, a person can roll again modifications to a venture to a particular cut-off date earlier than an accident occurred.
Masad demonstrated this throughout a Zoom name on Monday by instructing the Replit agent to destroy the contents of a database after which clicking on the software’s restore operate. Nevertheless, this operate is just not one thing a person can at present entry via the agent. They will need to have sufficient data of Replit’s options to find and use it.
Arguably, the principle drawback was that Lemkin and the agent had been successfully engaged on stay code, which meant that modifications had been instantly mirrored within the knowledge and efficiency of the stay net app. In customary software program improvement apply, new software program is constructed and examined inside a safe take a look at surroundings, usually referred to as a “sandbox,” and solely pushed stay as soon as all the pieces works as anticipated. That’s not how Replit features, at the least not right now.
“The issue I believe what we come clean with is correct now the database within the improvement surroundings is similar because the one the place you deploy it and go to manufacturing,” Masad says. “And so when the agent does one thing in a improvement surroundings it’s linked to manufacturing.” Masad provides that Replit responded to Lemkin’s state of affairs by working via the weekend to create a partition between a sandboxed improvement surroundings and the manufacturing surroundings. The corporate is within the technique of rolling out this new function now, he says.
“This stuff usually begin out with a whole lot of tough edges”
Masad mentioned on X that Replit will refund Lemkin his subscription payment for the difficulty and can conduct a postmortem on the incident.
We noticed Jason’s submit. @Replit agent in improvement deleted knowledge from the manufacturing database. Unacceptable and may by no means be attainable.
– Working across the weekend, we began rolling out computerized DB dev/prod separation to forestall this categorically. Staging environments in… pic.twitter.com/oMvupLDake
— Amjad Masad (@amasad) July 20, 2025
The episode could reveal one thing concerning the evolution of AI coding instruments, how they’re used, and what customers anticipate from them. Firms that create coding brokers, together with their supporters, usually boast concerning the spectacular outcomes these instruments can obtain via vibe coding.
However these glowing testimonials can overstate the advantages for common customers with various ranges of improvement expertise. (Lemkin didn’t instantly reply to Quick Firm‘s request for remark, however he did reply enthusiastically to Masad’s submit on X: “Thanks,” he wrote. “Actually admire you and all the assistance from the staff!)
Some AI coding instruments are useful for shortly constructing the entrance finish of an software however are much less succesful on the subject of forming and testing the backend knowledge connections that make an app purposeful. One developer at a big monetary providers firm mentioned vibe coding instruments usually fall brief on the subject of rigorously testing new options, in addition to testing each line of code within the bigger codebase that should be adjusted to accommodate the addition of an AI-built function.
Masad says that though Replit can free customers from the tedious syntax of coding, they nonetheless have to assume like builders. “You shouldn’t simply ask the agent for all the pieces,” he says. “You could be resourceful.”
He additionally acknowledges that coding instruments themselves should play a task in selling a developer mindset. “I don’t wish to absolve ourselves from duty, it’s incumbent on us as platforms to floor this data and to make it secure by default.”
And given the actual progress in reasoning, performance, and user-friendliness that AI coding instruments have remodeled the previous couple of years, it’s affordable to anticipate continued enhancements over the subsequent 12 months or two. “This stuff usually begin out with a whole lot of tough edges,” Masad says. “I believe the historical past of expertise has been that you need to be a bit of extra forgiving early on.”
