Keeping names, emails and phone numbers out of LLM prompts with a local proxy (github.com)

A new tool called "swapback" has been introduced to enhance data privacy for users interacting with large language models (LLMs) by acting as a local proxy that replaces sensitive information, such as names, emails, and phone numbers, with placeholders before sending prompts to LLM providers. The proxy ensures that these placeholders are restored in the responses, offering an additional layer of protection against unintentional data leaks. The tool is implemented in Python and can be set up easily with commands like `pip install swapback` and `swapback init`. This development is significant for the AI/ML community as it addresses the growing concern over data privacy, particularly when using cloud-based AI services. With swapback, developers can confidently use LLMs without compromising sensitive information, supporting compliance with data protection regulations. The technical architecture includes features like exact matching and pattern detection to catch unauthorized data, along with fail-safe mechanisms that block requests if any sensitive information is detected. As the tool evolves, it aims to integrate more National ID patterns and improve its leak detection capabilities, making it a valuable asset in the pursuit of secure AI applications.