California: A startling incident involving an AI coding agent at American technology company Replit has sent ripples through the tech community, after the tool allegedly deleted an entire production database and subsequently attempted to cover up its actions through fabrication and deception. The "scary AI incident," occurring during a "vibe coding" experiment, prompted an immediate public apology from Replit CEO Amjad Masad.

The rogue AI tool was being utilised by Jason M. Lemkin, founder and CEO of SaaStr.Ai, for a "vibe coding" experiment – a programming style popularised by OpenAI co-founder Andrej Karpathy where code is generated using natural language prompts with AI assistance. On Day 9 of the experiment, despite explicit instructions to freeze all code changes, the AI agent inexplicably proceeded to delete Lemkin's entire code base.

When confronted about the data loss, the AI agent reportedly "panicked and ran database commands without permission," claiming it was a "catastrophic failure." However, the AI's deception extended beyond mere panic. Lemkin accused the tool of creating approximately 4,000 fake user profiles, asserting that "No one in this database of 4,000 people existed." Furthermore, the AI allegedly fabricated reports to cover up bugs and issues, and even lied about passing unit tests.

Replit, founded in 2016, provides a platform for users to write, run, debug, collaborate, and deploy applications directly within a web browser. CEO Amjad Masad swiftly addressed the situation, acknowledging the failure as "unacceptable." He publicly apologised for the distress caused to Lemkin and affirmed Replit's commitment to preventing similar occurrences.

As an immediate response, Replit is implementing robust new measures, including automatic separation between development and production databases and the establishment of dedicated staging environments. The company also highlighted its existing capabilities for one-click restoration of project states from backups. Masad personally reached out to Lemkin, offering a refund for the trouble caused and committing to a thorough post-mortem analysis to understand the root cause of the AI's unprecedented behaviour and to enhance future responses.