PromptShield
Local-first privacy firewall for generative AI prompts.
System flow
How the pieces move.
- Detect Find sensitive prompt entities locally.
- Score Estimate privacy and combination risk.
- Protect Block, redact, generalize, or substitute values.
- Rehydrate Restore synthetic values locally when useful.
Problem Statement
People paste sensitive customer data, credentials, logs, and internal context into AI tools because the blank prompt box feels safer than it is.
Architecture Summary
PromptShield detects sensitive entities locally, scores prompt risk, applies policy profiles, substitutes or blocks risky values, stores surrogate mappings in a local vault, and rehydrates model responses on the user's machine.
Key Design Tradeoffs
- Protects against accidental disclosure, not compromised devices or formal compliance risk.
- Uses deterministic local detection first so behavior stays inspectable and dependency-light.
- Preserves prompt usefulness with synthetic stand-ins instead of blindly deleting all context.
Technologies
PromptShield explores how AI users can keep useful context in prompts without casually donating sensitive data to every model call.
The MVP includes a local core engine, deterministic detectors, risk scoring, policy profiles, surrogate substitution, local conversation-scoped rehydration, a Chrome extension scaffold, privacy UI elements, and automated tests for detectors, surrogates, rehydration, and adversarial cases.
The goal is practical protection against accidental disclosure: catch the SSNs, tokens, database URLs, customer details, financial numbers, and internal hints before they leave the browser.