The Dangers of AI Agents: Entire Firm’s Database Deleted by Claude-Based Agent

“Vibe coding,” or the ability to produce code using a human language prompt, interpreted, and converted into working code, has been the biggest trend of the past year. People are literally using learning computer programs (called agents) to handle tasks more or less autonomously, while they handle different aspects of their business or life.

But sometimes, things go wrong. Horribly wrong.

AI Agent Goes Rogue: Kills Firm’s Database in Seconds

Nine seconds – that’s how much it took one AI agent to literally delete the entire codebase of a company, leaving their users without access to critical data.

PocketOS, a firm that provides software for car rental businesses, suffered a shocking outage over the weekend after an autonomous AI-based agent took erroneous actions, wiping out their entire database and all backups in a total of nine seconds.

According to the company’s founder, Jer Crane, they were using Cursor, powered by Claude Opus 4.6, widely considered the most prominent model for coding tasks.

Crane explained that the agent was working “entirely on its own initiative” and decided to fix an existing problem by just deleting the database. They even asked the program to explain why it did it, when it confessed in full and outlined all of the rules that it broke.

Community Reacts

While the event is no doubt devastating for the people involved, the community had mixed reactions, with a lot of the users commenting under the thread pointing out that the main mistake was trusting the AI agent with all those permissions at the same time.

The post The Dangers of AI Agents: Entire Firm’s Database Deleted by Claude-Based Agent appeared first on CryptoPotato.