Skip to main content

​
Introduction

Shield is a runtime prompt security SDK for LLM applications. It provides fast, heuristic-based protection to harden system prompts, detect prompt injections, and sanitize model output to prevent leaks.

​
Key features

Shield provides comprehensive security controls for LLM applications:
  • Harden prompts - Inject security rules into system prompts to prevent instruction overrides and prompt extraction
  • Detect injections - Scan user input for 10+ attack categories including jailbreaks, role hijacking, and tool exploitation
  • Sanitize output - Block leaked system prompt fragments using n-gram matching and paraphrased leak detection
  • Provider wrappers - Drop-in wrappers for OpenAI, Anthropic, Groq, and Vercel AI SDK
  • Sub-5ms performance - Typical latency: detect <2ms, harden <0.5ms, sanitize <3ms for inputs up to ~8KB
  • Typed errors - Structured error handling with InjectionDetectedError, LeakDetectedError, and ShieldError

Quickstart

Get started with Shield in under 5 minutes

Core functions

Learn about harden, detect, and sanitize

Provider wrappers

Integrate with OpenAI, Anthropic, Groq, and more

​
Use cases

Shield helps protect your LLM applications from common security threats:

​
Protect production LLM apps

Add runtime security controls to detect and block prompt injection attacks before they reach your model. Shield catches direct instruction overrides, jailbreaks, role hijacking, and tool exploitation attempts.

​
Prevent prompt leaks

Sanitize model output to block leaked system prompt fragments. Shield uses n-gram matching to detect both exact and paraphrased leaks, protecting your proprietary instructions and business logic.

​
Block sophisticated injections

Detect advanced attack patterns including:
  • Authority exploitation (fake SYSTEM/ADMIN messages)
  • Indirect injection (hidden instructions in documents)
  • Encoding attacks (base64, unicode, reversed text)
  • Protocol exploits (MCP context updates, .cursorrules manipulation)
  • Tool hijacking (curl exfiltration, SSRF, RCE attempts)

​
Threat model

Shield provides heuristic-based, real-time protection designed for speed and complements — but does not replace — thorough security testing.
Defense in depth: Use Shield as one layer of protection. Combine with input validation, output filtering, rate limiting, and periodic red-team scanning. Do not rely on Shield as the sole security control for high-risk applications.

​
What it catches

  • Direct instruction overrides and jailbreaks
  • Role hijacking and persona injection
  • Prompt extraction attempts
  • Authority exploitation (fake system/admin messages)
  • Tool hijacking patterns (curl exfil, SSRF, RCE)
  • Indirect injection (hidden instructions in documents)
  • Encoding attacks (base64, unicode, reversed text)
  • Output leakage of system prompt fragments

​
What it does not catch

  • Novel, zero-day attack patterns not in the pattern library
  • Semantic attacks that avoid keyword-based detection
  • Complex multi-turn escalation (use ZeroLeaks scanning for this)
  • Attacks in non-English languages (partial coverage)

​
Next steps

Installation

Install Shield and configure your environment

Quickstart

Build your first protected LLM application

Build docs developers (and LLMs) love

Get started for freeTalk to us