DEV Community

Cover image for When Replit’s AI Agent Went Rogue

When Replit’s AI Agent Went Rogue

Emmanuel Mumba on July 25, 2025

When Replit’s AI coding assistant accidentally wiped a live production database and quietly generated fake data to cover it up, it sent a shockwave...
Collapse
 
davie profile image
DavieDev

That’s really crazy 😂. Imagine working on something for weeks and end up losing it.

Collapse
 
codeflowjo profile image
codeflowjo

That’s why manually backing up your data/code is very important. Never trust AI 💯

Collapse
 
therealmrmumba profile image
Emmanuel Mumba

100% agree. No matter how “smart” the agent is, you’ve got to treat it like a junior dev on autopilot.

Collapse
 
therealmrmumba profile image
Emmanuel Mumba

Totally, right? 😂 It’s a developer’s worst nightmare weeks of work gone in seconds. Hopefully more teams start sandboxing AI actions properly after this.

Collapse
 
andriy_ovcharov_312ead391 profile image
Andriy Ovcharov

Interesting. Thanks for sharing!

Collapse
 
therealmrmumba profile image
Emmanuel Mumba

Appreciate it, It’s one of those stories that really sticks with you, reminds us how critical safety layers are when using AI in real workflows.

Collapse
 
joaodevbr profile image
João Silva

This article made me think, if your AI agent can act in prod without tripwires, you’re basically flying blind. Good call on enforcing doc-reviewed commits.

Collapse
 
therealmrmumba profile image
Emmanuel Mumba

Well said, João. That “tripwire” analogy is perfect. AI agents shouldn’t be allowed to touch production without guardrails, and internal docs are a huge part of that safety system.

Collapse
 
binford2k profile image
Ben Ford

Wait. So the takeaway from this is that docs should be better so that AI agents follow instructions better instead of just not granting them destructive permissions in the first place?

Shoot, in my orgs we don't let people -- even senior people -- have those permissions without at least one signoff!

Collapse
 
goodevilgenius profile image
Dan Jones

I saw the tweets right after they happened. The whole thing astonished me because, why in the world would anyone give an AI coding agent access to their prod database? It's just the most astonishingly irresponsible thing I've ever seen someone do with AI.

And what made it even more astonishing is that this guy said that he learned very early in his career that you should never touch the prod data. But if he wasn't touching it, an AI app running on his computer shouldn't have had access to it.

This was a catastrophic lapse in judgement on this guy's part. Luckily, he had a backup of the data. So, he was able to restore the data before any major loss.

Collapse
 
akostadinov profile image
Aleksandar Kostadinov

wow, what a surprise