Preprint tests reversible personal-data substitution for LLM prompts

Researchers report in a preprint that Veil-then-Reveal replaces personal information on a user's device before sending prompts to remote language models, then restores original values in responses. Tests on three datasets containing 11,322 annotated personal-information instances achieved 97.9% detection recall and 95.2% protection of detected values. The framework substitutes realistic values of the same type, using encryption that preserves formats for structured information and a keyed dictionary and secure vault for names, locations and organizations. Session-specific mappings prevent links across sessions. The researchers report 100% restoration of echoed substitutes, but the detection and protection results leave gaps in coverage.