Rising Copilot spending led to an experiment with a MITM proxy

When GitHub Copilot credits started running out noticeably faster, the author decided not to stop at guesswork and to look at how the tool works inside. For that he chose VS Code and put Copilot behind a MITM proxy — an intermediary through which an application's network traffic can be observed.

That is an important turn in the approach to AI services. Instead of the question "how well does it write code" comes another: which requests are sent, how does the application talk to the server, and what exactly might affect the consumption of limits.

But there is a catch: the available excerpt of the material describes only the setup of the experiment. The specific observations, technical details and conclusions of the investigation are not given here, so pretending the cause of the accelerated credit burn has already been established would be wrong.

The practical value of the approach is clear all the same: a proxy turns an opaque desktop application into an object for observation. Such an experiment does demand care, though — intercepting network traffic can touch sensitive data, and without a full analysis of the results it remains a way to gather facts rather than a ready explanation.

When an AI tool starts burning through limits faster, do you first look for the cause in your own requests, or are you ready to dig into what is happening inside the application? Source: All articles / Artificial Intelligence / Habr