taintguard

⚠️ Experimental – Not for production

This package is a reference implementation / idea starter for span-based taint tracking and guardrails around LLM prompts. It is not a drop-in security control and comes without guarantees. Use it for exploration, prototypes, or as a basis for your own hardened implementation.
For any production use, perform your own threat modeling, add comprehensive tests, and combine with defense-in-depth controls (rate limiting, moderation, output validation, allowlists, human review, etc.).

Span-based taint tracking for LLM prompts (TypeScript-first). Keep untrusted text from hijacking system instructions or tools.

Why taintguard?

Prompt injection thrives on mixing instructions with data. Regex-only filters are brittle.

Features

🧱 Prompt fencing: XML-like … by default
🔎 Preflight detection: jailbreak/role-hijack/developer mode/tool-trigger patterns
🌍 Language packs: English & German rules out of the box
🧪 Property tests: adversarial fuzzing (zero-width, homoglyphs, spacing) with fast-check

Quick Start

pnpm i
pnpm build
pnpm test

Branch Code Coverage: 77.77%