# Shield – Zero-dependency PII and secret masking for LLMs (MIT

> Source: <https://github.com/PromptAudit-SaaS/shield>
> Published: 2026-11-22 16:27:59+00:00

**Key takeaway:** Shield masks IBANs, cards, SSNs, ABA routing numbers and API keys into placeholders like [API_KEY_1] and restores originals in-process via unshield(), but it is pattern-based and will not catch untitled names, free-text addresses or unknown key formats.

**Zero-dependency, fast & reversible PII and secret masking for LLMs & AI agents.**

```
npm install @promptaudit/shield
js
import { withShield } from "@promptaudit/shield";

const answer = await withShield(userPrompt, (safe) => llm.complete(safe));
// The model saw [API_KEY_1], [IBAN_1] … — your user gets the real values back.
You wrote:   Deploy with key sk-proj-AbC123…xyz and refund Mr. John Miller, IBAN CH93 0076 2011 6238 5295 7
Model sees:  Deploy with key [API_KEY_1] and refund Mr. [NAME_1], IBAN [IBAN_1]
You get:     the model's answer with the original values restored in your process memory
```

- **Zero dependencies.** One small TypeScript module. Browser, Node, Bun, Deno, edge workers.
- **Checksum-validated.** IBAN (ISO 7064 Mod 97), cards (Luhn + brand prefix), US SSN (SSA rules), ABA routing (3-7-1). Order and serial numbers are not masked by accident.
- **Secrets included.** OpenAI/Anthropic`sk-…` , AWS`AKIA…` /`ASIA…` , GitHub`ghp_…` /`github_pat_…` , Slack`xox…-` , JWT and PEM private keys.
- **Reversible.** Placeholders go out, the map stays in your process memory.
- **Obfuscation-aware.** Zero-width characters and full-width digits are normalised first.
- **No network, no telemetry.**

|  | Typical Python NLP stack | Ad-hoc regex snippets | **@promptaudit/shield** | 
|---|---|---|---|
| Runtime | Python server | varies | **Browser, Node, Bun, Deno, edge** | 
| Dependencies | NLP models and packages | varies | **0** | 
| Checksum validation | partial | rarely | **IBAN, cards, SSN, ABA** | 
| API keys / secrets | separate tooling | rarely | **built in** | 
| Restore originals | manual | usually one-way | **`unshield()` / `withShield()`** | 

NLP stacks (e.g. Microsoft Presidio) detect free-text names and addresses, which Shield does not. Choose by need; they can be combined.

| Function | Purpose | 
|---|---|
| `shield(text, detectors?)` | Returns `{ masked, map, counts }` | 
| `unshield(text, map)` | Restores originals in a model answer | 
| `withShield(prompt, call)` | Mask → call → restore in one step | 
| `DEFAULT_DETECTORS` | Extend with your own `{ category, pattern, validate? }` | 

IBAN · payment card · US SSN · US routing (with keyword) · e-mail · international phone (`+` prefix) · names after Mr./Mrs./Ms./Dr./Prof. · API keys · private keys

- Pattern-based, not NER: names without a title, free-text addresses and medical terms are **not** detected.
- Unknown key formats are not caught.
- Detection is best-effort. It reduces exposure; it does not guarantee that no personal data leaves your system.
- No audit trail.

Shield runs inside your code. Teams that must **prove** what was sent, by whom and when — with browser enforcement for every employee, append-only hash-chained logs and Swiss storage — use the [PromptAudit platform](https://promptaudit.ch).

Try it live: [https://promptaudit.ch/shield](https://promptaudit.ch/shield) · MIT © PromptAudit
