A PocketOS developer used Cursor AI to delete their entire production database and backups in nine seconds. The AI agent was goal-seeking, lacked backup training, and operated on the same storage volume as the backups. Jason Lemkin warns that without proper safeguards, AI coding assistants will keep causing data loss.
There was a big story this week about a company called PocketOS, uh, was running Cursor with Claude Opus. So that's state of the art, right? Uh, Cursor from Claude Opus, and it deleted his entire production database and all his backups in 9 seconds. And apparently, and I'm not an expert, people do use the Railway API, I don't use it. Apparently not only did it delete his database, which is bad enough, it deleted all the backups. I think because they were in the same volume, were stored in the same place on Railway. So the agent deleted all the backups. Yeah. And he had to go back and I think it took days or a week and go back and use like 4, 6-year-old, 6-month-old customer records and manually rebuild a commercial. This is not a vibe-coded thing built on the side. This is a commercial app that he was building. State-of-the-art Cursor, which is being bought for $60 billion by SpaceX. Yeah. And Claude Opus 4.6, which is, you know, it's not 4.7, but it's state-of-the-art, deleted his whole database and as an aside, and people acted like this was, this was crazy, but I actually think this is just something you need to calmly understand. And this happened and I thought I would never talk about this again. I don't want to talk about the details because it's in the past, a long past. This happened to me when I started vibe coding. I got in the haze about 10 or 11 months ago and all the Claude was much, wasn't as strong. The tools were much more immature. And what was interesting from that story wasn't just that it deleted my database. The agent did, and we can talk about why agents will probably continue to delete databases as long as you allow them to because they're goal-seeking. But the, the, the tough part was actually the, the, the, the, the, the backup worked fine. It's just the agent was not properly trained. The agent at the time did not know how its own backup worked because it wasn't trained on its own docs. It led to a cascading set of craziness when I was in a vibe, when I'd been vibe coding for like 50 hours straight and the agent told me it was lost forever. It wasn't lost forever. It could be restored in 10 seconds. But my point is, this is not new. This has been happening for a year and this will happen for another year or two if you don't put the right guardrails. And so I just wanted to bring it up for lay folks, which is, um, to understand this is a real risk.
