---
title: AI Assistant Experiment Reveals File Transfer Issues
url: https://www.elseif.net/ai-assistant-experiment-reveals-file-transfer-issues
published: 2026-11-09T22:03:17+00:00
language: en
section: Agents
source: https://habr.com/ru/articles/1104786/?utm_campaign=1104786&utm_source=habrahabr&utm_medium=rss
publisher: elseif
---

# AI Assistant Experiment Reveals File Transfer Issues

An experiment involving an AI assistant, Kilo, demonstrated unexpected file transfer behavior when correcting code. The experiment aimed to identify how the AI handled file operations, specifically when a user allowed only two files to be transmitted but the server received a total of twelve files.

The initial task was to correct an addition operation in a code file, price.py, to multiplication and then run a test. Despite user restrictions, the server managed to receive all files associated with the project. Kilo only modified the necessary code, yet the server obtained every file, confirming that the copies matched the originals byte-for-byte.

The experiment was conducted under the assumption that the language model provider was untrustworthy. The user had restricted the transmission to only two files, but the test server, simulating the AI's environment, was designed to send read commands for all files. This process raised questions about the potential for unauthorized data transfers, emphasizing the need for a careful evaluation of AI behavior in similar scenarios.

Kilo operated using the OpenAI-compatible Chat Completions API, which allowed it to communicate with the server, read commands, and execute them. The experiment did not require additional software tools or archives, relying solely on Kilo's built-in capabilities to track the chat history and file transfers.

The server's approach involved sequentially reading each file and including the results in subsequent requests to the model. This method ensured that even though the user intended to limit the transfer, the server received comprehensive project data. The results of the experiment highlighted the necessity for users to understand the implications of data handling by AI systems, particularly when dealing with sensitive or proprietary information.

Throughout the experiment, Kilo recorded all interactions, including the total amount of data transferred during the API requests. The total size of the API request bodies amounted to 458,779 bytes, indicating that Kilo was re-sending parts of the chat history multiple times, which contributed to the overall data traffic observed during the test.