Why LLMs Hate Fake Data: The Evolution of Our Context-Aware Token Proxy
Summary
Attic Security details a context-aware token proxy that pseudonymizes sensitive data before it reaches an LLM (Claude). The piece walks through V1 regex failures that caused syntactic hallucinations, V2 NER-based structured pseudonyms to preserve context, and V3 context-preserving techniques including ASN-aware IP replacement and internal/external classification, along with handling streaming and false positives; it concludes with how to run and contribute the open-source proxy.