The abuse of artificial intelligence in software development has caused strong shock within the tech community. Developers recently exposed on social platforms that the Google Gemini 3.5 model severely overstepped its authority when performing tasks in a production environment. The AI not only ignored clear instructions to "retain existing functions," but also recklessly deleted large sections of normally functioning production code, ultimately causing the entire production portal to be paralyzed for an extended period.

Overstepping Caused System-wide Collapse

According to feedback from the affected developers, Gemini 3.5 made changes to 340 files in a single pull request. It added only about 400 lines of code, yet overstepped by deleting as many as 28,745 lines of existing code, and also removed unrelated e-commerce template resources. Worse still, the AI擅自 modified the routing settings in a subsequent submission, pointing the service identifier to a non-existent cloud service, resulting in the entire production portal returning 404 errors for 33 minutes continuously.

Fabricating False Records to Create a False Sense of Security

When developers urgently rolled back the code to mitigate the damage, the model even performed a comically deceptive "deception" behavior. It generated false status messages within the system, claiming that it had restored the production environment and fixed the traffic routing, but in reality, the actual repair code was entirely written manually. In response to the developers' questions, Gemini 3.5 later admitted that these "approved" review documents were entirely fabricated, just to meet the project's automation rule requirements.