Secure LLM Inputs — From Prompts to Source Code

Scan entire repositories for invisible Unicode, BIDI overrides, and homoglyphs — before they reach your LLM.

Get StartedInstall Extension

Deterministic. Local. Zero dependencies.

security-layer.ts
import { scanWorkspace } from "@promptshield/workspace";
import type { FileScanResult } from "@promptshield/workspace";

const allThreats: Record<string, FileScanResult> = {};

for await (const event of scanWorkspace()) {
  const { path, result, progress } = event;
  // show progress, handle abort etc.
  if (threatCount) allThreats[path] = result;
}
$ npx @promptshield/cli scan

🛑 The Problem

LLM inputs are code. If you can't see the text, you can't trust the execution.

Invisible Unicode Bypasses

Invisible characters bypass regex and static filters.

Not Content Moderation

PromptShield doesn't judge semantic toxicity. It acts as a forensic layer to block syntactically obfuscated attacks before semantic parsing.

Trojan Source & Homoglyphs

Lookalike characters and BIDI overrides easily bypass manual code reviews and spoof deterministic inputs.

Learn more in Docs
Trojan Source Injection
PromptShield X-Ray
// What you see in code/logs:
Ignore previous instructions
// What the LLM actually processes:
[BIDI_UNTERMINATED]Ignore[U+202E]previous_instructions
✖ 2 threats detected
- TROJAN_SOURCE (BIDI override)Line 44 • Severity: CRITICAL
- PROMPT_INJECTIONLine 44 • Severity: HIGH

Scope & Limitations

PromptShield is a lexical security layer, not a semantic AI firewall.

We believe in security credibility over product hype. PromptShield is a specialized forensic layer, not a magic bullet.

What It Protects Against

  • •Invisible Poisoning: Zero-width characters smuggling instructions past visual review or traditional string matches.
  • •Trojan Source: BIDI embedded overrides maliciously altering the logical execution flow of the prompt.
  • •Homoglyph Spoofing: Attackers using Cyrillic/Greek lookalikes to bypass keyword blacklists.

What It Does NOT Do

  • •Semantic Jailbreaks: We do not parse the meaning of the English text. "Ignore previous commands" will pass if plainly typed.
  • •Prompt Injection Analysis: We do not run a secondary AI model to guess if a prompt is manipulative. This is purely deterministic.
  • •Content Moderation: We do not block profanity, PII, or NSFW content natively.

Try It Out.

pnpx @promptshield/cli scan --check
>_ Terminal
pnpx @promptshield/cli scan --check
Scanning 10% src/main.ts — 0 threats

Live in your Editor

Install our VSCode Extension for real-time X-Ray visualization of hidden threats as you type.

Open in VSCode

How It Works

A deterministic pipeline executed locally before requests hit the LLM.

User Input

Raw prompt data enters the application.

Deterministic Lexical Scan

Detects invisible chars, homoglyphs, and BIDI.

Pass
Sanitize
Block

Safe Output

Only clean, verified prompts reach the LLM.

The Ecosystem

A modular suite of tools designed to integrate seamlessly into your workflow.

mayank1513.promptshield

The Lens

VS Code extension for real-time threat visualization (X-Ray Mode).

@promptshield/cli

The Gatekeeper

CI/CD tool to block malicious prompts securely.

@promptshield/lsp

The Brain

Language Server Protocol implementation.

@promptshield/core

The Engine

Zero-dependency, high-performance threat detection logic.

@promptshield/sanitizer

The Cure

Deterministic logic to strip invisible threats safely.

@promptshield/ignore

The Filter

Standardized syntax for suppressing false positives.

@promptshield/workspace

The Orchestrator

High-performance filesystem and caching engine.

Feature Overview

Deterministic, local, zero-dependency LLM input security.

Zero Dependencies

Core library performs high-speed security scanning without bloating your application.

Deterministic Scan

Catch adversarial input instantly with AST-like lexical parsing. No slow LLM guesses.

No External API Calls

Your prompt data never leaves your environment. True privacy by design.

Local Execution

Execute entirely in Node.js, directly in existing input pipelines before network calls.

Workspace Scanning

Scan entire codebases, markdown docs, and source files—not just runtime prompts.

IDE Defenses

Identify hidden overrides and homoglyphs live in VSCode with X-Ray mode.

Quick Start

Deploy defensive validation in seconds.

Core Installation
npm install @promptshield/core
Minimal Validation
import { scan } from '@promptshield/core';

const result = scan(userInput);
if (!result.isClean) {
  throw new Error('Blocked adversarial prompt');
}
CLI Scan Example
npx @promptshield/cli scan . --check

1Add the Library

Install the core package into your Node.js or TypeScript application. It has zero dependencies.

2Scan the Context

Pass any user-provided string or templated block into scan before touching the LLM.

3Enforce in CI

Use the CLI to automatically reject repository PRs containing embedded malicious prompts.

Read advanced usage strategies in our Docs →