A new open-source tool called Rampart aims to prevent personal information from leaving a user's device when interacting with chatbots or other AI services. Developed by a team focused on privacy-preserving technologies, Rampart functions as an on-device filter that redacts personally identifiable information (PII) before any message is transmitted. The system operates entirely within the browser, eliminating the need for remote servers or external processing.

When users type messages into AI interfaces, sensitive details such as names, addresses, Social Security numbers, or financial data can be inadvertently exposed. Rampart addresses this by intercepting text at the moment of composition and removing or masking such information locally. The filtering process uses two components: a rule-based system that identifies structured data like phone numbers and email addresses through regular expressions, and a lightweight language model called MiniLM that detects contextual mentions of names and locations.

The project was motivated by concerns about data handling practices in AI systems. Many existing PII removal tools either require uploading user input to external servers or depend on large models that are impractical for widespread browser use. Rampart's developers note that AI vendors often cannot guarantee privacy, and model updates may introduce unforeseen risks. Additionally, large privacy filters can exceed 2.8GB in size, creating significant download barriers for users with limited bandwidth.

Rampart's design emphasizes efficiency and accessibility. The entire pipeline, including the MiniLM model and tokenizer, occupies approximately 14.7MB, enabling rapid processing with latency under 4 milliseconds in WebGPU-enabled browsers. In benchmark testing using a dataset of 30,000 anonymized examples across seven languages, Rampart achieved a 98.4% success rate in identifying PII. Performance varied across comparison models, with some requiring significantly larger footprints or relying on cloud-based processing.

The system currently supports English, Spanish, French, German, Italian, Portuguese, and Dutch. It is released as an alpha-stage tool under an open-source license, available through Hugging Face and via an NPM package for developers. The team also published a whitepaper detailing the architecture and evaluation results.

Rampart does not claim to solve all privacy challenges in AI interactions. Rather, it represents an initial step toward on-device protection of sensitive data. The developers emphasize that it should be considered part of a broader strategy for managing personal information in digital environments. Users and developers interested in testing or integrating the tool can access the code and documentation through the provided platforms.