PrivacyScrubber
This article has been nominated for deletion. You are welcome to participate in the deletion discussion, which will decide whether to keep it. This discussion may also result in the article being merged, redirected, or draftified. |
| PrivacyScrubber | |
|---|---|
| Original author | Ilya Sibiryakov |
| Developer | Brand Me Web |
| Operating system | Cross-platform (Browser extension, CLI) |
| Type | Data loss prevention software, Computer security |
| Website | privacyscrubber |
PrivacyScrubber is a client-side data sanitization tool and security gateway designed to redact Personally Identifiable Information (PII) and Protected Health Information (PHI) before transmission to cloud-based Large Language Models (LLMs). It employs a Zero-Trust Data Sanitization (ZTDS) architecture, which aims to process sensitive data within the local computing environment.[1]
Architecture and technology
[edit]PrivacyScrubber operates entirely offline. The core tokenization engine intercepts sensitive entities (such as names, Social Security numbers, and credit card PANs) and replaces them with synthetic tokens (e.g., [NAME_1]).
The state mapping linking the original data to the tokens is encrypted in the browser's volatile memory using the XChaCha20-Poly1305 authenticated encryption algorithm, with keys derived via Argon2id.[2] Temporary extraction buffers are overwritten at the byte level prior to memory garbage collection.
Integration
[edit]PrivacyScrubber functions across multiple environments:
- Web Application & Browser Extension: Integrates into the Document Object Model (DOM) to sanitize text inputs for web-based AI interfaces.
- Model Context Protocol (MCP): Functions as a local MCP server, creating a local context window for autonomous AI developer tools to manage the exposure of environment variables and source code.[3]
Compliance framework
[edit]The tool is designed to address several regulatory frameworks by shifting the sanitization workload to the endpoint:
- HIPAA & HITECH: Redacting PHI identifiers.
- FERPA: Masking student identities.
- PCI-DSS & GLBA: Securing consumer financial records.
- GDPR & CPRA: Supporting data minimization and the Right to be forgotten.
References
[edit]- ↑ Sibiryakov, I. (2026). "Zero-Trust Data Sanitization: A Client-Side Architecture for Secure LLM Integration". Zenodo. Zenodo. doi:10.5281/zenodo.21796958.
- ↑ Biryukov, A.; Dinu, D.; Khovratovich, D. (2016). "Argon2: the memory-hard function for password hashing". 2016 IEEE European Symposium on Security and Privacy (EuroS&P). IEEE. pp. 292–302.
- ↑ "Model Context Protocol (MCP) Specification". Anthropic. 2024. Retrieved 2026-08-04.