Show HN: Pseudonymizing sensitive data for LLMs without losing context (atticsecurity.com)

A new tool called Ghost Analyst has been developed to enhance incident triage in cybersecurity by utilizing Anthropic’s Claude language model while ensuring sensitive data remains secure. The innovation lies in a Data Loss Prevention layer that pseudonymizes client data without compromising the reasoning capabilities of the LLM. Initially, basic regex solutions faced significant challenges, including "hallucination" issues where the model created fictional characters leading to failed queries. This prompted an evolution to a context-aware proxy, which maintains the critical contextual information while obfuscating sensitive data. The final version (V3) of the pseudonymization process employs an advanced token proxy that intelligently replaces sensitive details with context-preserving pseudonyms. By leveraging techniques such as ASN-aware IP replacement and internal/partner/external categorization, the model can accurately assess security incidents without ever processing real user data. With a focus on minimizing false positives in detection and providing seamless integration with other models, this breakthrough offers a robust solution for environments needing to balance data privacy with AI efficiency. The project is open-sourced, allowing the wider AI/ML community to adopt or adapt the technology for various applications beyond cybersecurity.