Uploaded a Wrong File to an AI — Fired. How Data Security Works in AI

The article by Daria Lushkina from the RuNet Rating examines data security when using neural networks. It notes a study that found employees of 150 Russian companies uploaded 30 times more corporate data to public AI services in 2025 than a year earlier, and 60% of companies have no AI usage rules. In July, a top manager of a Moscow engineering company was fired for uploading work documents to DeepSeek, and the court sided with the employer. The article details the data processing pipeline: files are parsed, text extracted, split into fragments, and converted into embeddings, leaving traces in logs and storage. Data can be vulnerable at storage, logs, third-party chains, and training. Deletion is complex: removing a dialogue does not delete all copies, and if data was used for training, its influence is irreversible. The article also covers machine unlearning as a research direction. It notes that human moderators may review dialogues, as in Google's Gemini, and that infrastructure partners like cloud providers have access. It mentions a 2023 DeepSeek data breach where over a million rows of chat logs and keys were exposed. Samsung engineers sent source code to ChatGPT, and a US cybersecurity agency head uploaded sensitive documents. The author advises caution and understanding data policies.