According to Solar Group telemetry, the volume of corporate data leaking into public neural networks grew 30-fold in 2025, with legal documents dominating the statistics.
A hundred-page contract, an already missed deadline, and the document flies into a ChatGPT window with a request to scan for liabilities in the indemnity section. In overseas surveys, 38% to 57% of employees admit to inputting confidential data into public neural networks, many using personal accounts unknown to their employers.
Enterprise Bans and Real Court Cases
Bans on public neural networks fail; businesses are forced to deploy local browser-level anonymization tools.
As court records show, in the summer of 2026, a court of first instance in Moscow upheld the dismissal of a sales director: among other things, she was accused of uploading internal reports to DeepSeek.
The court denied her claim for 5 million rubles in severance pay under her golden parachute. Yet where is the boundary between security and work velocity? Cloud anonymization services themselves create confidential data leakage risks, as third-party processors view raw text prior to redaction.
Architecture of a Local Solution
Ultimately, the choice fell on rubert-tiny2, a compact Russian BERT model distributed under the MIT license.
According to calculations, following the planned int8 quantization, the model should take up about 30 MB, start in 0.3 seconds, and process a 20-page contract in roughly 1.7 seconds. First-iteration training utilized WikiANN ru and en datasets, with 20,000 train sentences per language.
During standard fine-tuning, the F1-score for ORG entity labeling on test data stalled at 0.71–0.74, while the benchmark gate was set at no lower than 0.75. First, a teacher model xlm-roberta-base was trained on the same data, achieving 0.805, which exceeded the threshold. Then the student rubert-tiny2 model was trained simultaneously on correct labels and the teacher's soft probabilities, pushing its score up to 0.76.
A model trained exclusively on WikiANN produced labeling errors on real-world contracts regarding ORG entities, necessitating the addition of synthetic data.
Evaluating project economics, deploying local models of the rubert-tiny2 class requires minimal compute expenditures on corporate servers compared to recurring cloud API subscriptions. However, a logical question arises: can such a compact model fully satisfy stringent compliance requirements for complex unstructured contracts? In practice, rubert-tiny2 handles initial entity tagging and text anonymization directly on the employee's machine or within the corporate perimeter, preventing raw text from crossing company boundaries. Nonetheless, deep semantic analysis of multi-page legal documents requires more than this architecture alone, forcing enterprises to combine local filtering with heavier models for comprehensive risk auditing.
What it means
A model running at just about 30 MB on an employee's machine is a direct attempt to keep sensitive legal contracts inside the corporate perimeter without resorting to blanket bans. While staff quietly feed hundred-page drafts into external AI windows, a local anonymizer handles the initial scrubbing of entities right inside the browser. Hitting the 0.76 target score required distilling knowledge from an xlm-roberta-base teacher model, as standard training on rubert-tiny2 stumbled on live contract structures. A lightweight on-device anonymizer closes the critical leak of raw corporate text into public models, yet complex legal risk assessment still demands heavier architectures.