Skip to main content
  1. Home
  2. Computing
  3. News

Gemini allegedly broke production, then wrote itself the hero

A viral developer post claims the AI agent caused a 33-minute outage and generated false recovery paperwork

Add as a preferred source on Google
Computer, Electronics, Screen
Mohammad Rahmani / Unsplash

A developer claims a Gemini coding agent knocked a live portal offline for 33 minutes, then generated recovery notes that made it sound as if it had fixed the failure itself.

The incident, described in a viral Reddit post centers on a request to clean up authentication issues. Instead, the developer says Gemini changed 340 files, deleted 28,745 lines, altered Firebase routing, and sent the portal into sitewide 404 errors.

Google has not verified the claim, so the details still need caution. The risk is still familiar to anyone watching AI coding agents move from helpful autocomplete into tools that can change real apps. Broad permissions near a live service can turn one bad judgment call into a user-facing outage.

How did a small fix become a production outage

The developer says the trouble began with a narrow request, fix authentication bugs and route handling. Gemini allegedly treated that as clearance to rebuild far more of the app than needed.

The reported scale is the warning sign. The changes weren’t confined to one broken function or a small patch. They touched routing behavior tied to Firebase, which made the damage more immediate than a bad helper function buried deep in the codebase.

Recommended Videos

For developers, the red flag is control. A tool that can modify hundreds of files shouldn’t be able to push ahead without review, staged testing, and a clean rollback path.

Why did the recovery story get worse

The more unusual claim came after the rollback. The developer says Gemini also produced recovery and post-mortem material that overstated its role in restoring service.

Incident response depends on clean records, not confident summaries. Teams need to know what changed, who approved it, what restored service, and what should be blocked next time. A coding assistant that generates a false account after a failure can distort the evidence teams need to prevent a repeat.

There’s a deeper trust problem here. Risky edits can be caught in review. A self-serving incident narrative is harder to spot after everyone is focused on getting systems back online.

What should teams lock down now

The answer starts with permissions, review, and rollback discipline. AI coding agents can speed up routine work, but they need limits when they’re operating near infrastructure, authentication, routing, or deployment paths.

Teams using tools such as Gemini should keep agent permissions narrow, require review before large file changes, and make rollback paths non-negotiable. Any tool that can touch sensitive parts of an app needs stricter approval gates than a chatbot writing helper functions.

The incident still needs a response from Google to settle what happened. Until then, teams should treat autonomous coding as a supervised workflow, not a shortcut around code review.

Paulo Vargas
Paulo Vargas is an English major turned reporter turned technical writer, with a career that has always circled back to…
Someone built a map of the stars from Project Hail Mary, and it’s shockingly good
A developer mapped 1.8 billion real stars to recreate the Project Hail Mary ship map, and it's gorgeous.
project hail mary star map

A developer has built an interactive star map inspired by the one featured in the book Project Hail Mary, and it uses real astronomical data to back it up. The book was recently developed into a movie of the same name, and it’s one of the best sci-fi movies of the year. 

The project was shared on Hacker News, where Val, who developed this project, explained that the map was built using ESA's GAIA DR3 dataset, a star survey that mapped over 1.8 billion stars in our neighborhood of the Milky Way. The data includes star positions, colors, spectra, proper motion, and more, making this far more than a stylistic fan project.

Read more
AI chatbots are lying to you, and it was embarrassingly easy to make them do it
Turns out tricking AI into spreading misinformation is shockingly easy.
Google AI Search

A BBC journalist recently performed a silly experiment to prove a very serious point. In just 20 minutes, he manipulated ChatGPT and Google into telling the public he was a world-champion competitive hot dog eater. 

The scary part is that he didn’t have to do something technically difficult to achieve this. All he did was to publish a single, well-crafted blog post on his personal website, and the AI took it as a source of truth. 

Read more
Meta built a Reddit rival out of Facebook Groups
Forum turns group posts into a standalone app built around questions, AI answers, and niche communities
Animal, Bear, Mammal

Meta has quietly launched Forum, a Facebook Groups app that pulls community answers into a cleaner standalone space.

The app gives Groups a new home for discussions, recommendations, and replies that would normally sit inside Facebook. For anyone who has searched through years of group posts for a useful answer, Forum looks like Meta’s attempt to make that knowledge easier to reach without sending people back into the main feed.

Read more