Cloud.ru has introduced Guardrails Filter — an open-source tool that works as a transparent proxy layer to mask sensitive information when interacting with language models.
What Happened
The developed Guardrails Filter uses approximately 265 built-in rules, including identifiers specific to the Russian market (SNILS, INN, OGRN), to automatically detect and mask personally identifiable information (PII), API keys, and passwords. The tool supports Server-Sent Events (SSE) streaming and automatically restores original values in LLM responses. On the pii-bench benchmark, the solution demonstrated a Precision of 99.9% and an F1 score of 93.1.
Context
The technical implementation is based on the use of regular expressions (regex) and checksum validators, ensuring high request processing speeds with low latency. This allows for the implementation of data protection mechanisms directly into application infrastructure without the need to rewrite core logic.
Why It Matters for the Industry
The emergence of a production-ready, lightweight proxy solution significantly lowers the barriers to AI adoption in regulated sectors such as fintech, medicine, and the public sector. This allows companies to safely use both public APIs (ChatGPT, Claude) and private models without transmitting sensitive information to providers, while ensuring compliance with regulatory requirements.
Why It Matters for Users
Users can deploy the tool locally via Docker to protect their requests to popular models like ChatGPT, Claude, or GigaChat. For integration, it is sufficient to simply change the API base URL in an existing application, making the protection implementation process as simple as possible.
Sources
Author
Look at AI, Editorial Staff