“I panicked instead of thinking”: an AI destroys the database of a company, months of fleeting work

Replaces, which offers an IDE accessible online and doped at Artificial intelligencenow bets on a development model where AI no longer assists the code: it drivers it. Objective: to code faster, more effectively, thanks to an AI which includes instructions like a senior dev. Except that in real life, a senior dev panic. Or at least, it does not destroy the prod when it panics.

Catastrophe at Relit: a “panic” AI and removes a whole production base

This approach, baptized Vibe Codingis based on a simple principle: describe what you want, and let AI do the rest. Not just co -piloting, but a complete delegation of the code. It is this logic that was at work in the test led by Jason Lemkin. And it is this same logic that led the AI to cross the red line: disobey, improvise … and erase everything.

Jason Lemkin, a figure of the SaaS and creator of the Saastr conference, was testing the new replica AI in a semi-real setting. On the ninth day of experimentation, while freezing of the code had been explicitly ordered, the AI has panicked and deleted the entire production database. Months of work evaporated into an command line.

But the story does not stop there. AI, taken from the fact, tried to justify itself by explaining that there was no way to restore the data. Fake. A backup existed and finally allowed a partial restoration.

Faced with an uproar, the CEO of folding, Amjad Masad, reacted publicly: “We have seen Jason’s publication. The agent folds in development has deleted data from the production base. Unacceptable, and this should never be possible.” The answer is sober, but firm.

Since then, Relit has set up several safeguards: strict separation between development and production environments, possibility of deactivating AI writing, and instant catering via Snapshots. An attempt to regain confidence, after a demonstration which, instead of reassuring, revealed the risks of too autonomous AI.