Lakera Guard

What it is

A managed classifier for hostile input.

Lakera Guard is a real-time AI security API. In a single call it runs ML-based detection for prompt injection, jailbreak attempts, PII, and malicious links. Because it is model-based rather than a list of regexes, it catches semantic attacks: the phrasing that means "ignore your instructions" even when the exact words are new. It is provider-agnostic, so it sits in front of whichever model you call.

The one job

Catch injection before it lands.

Screen every piece of untrusted text, whether typed by the user or pulled from a document, and flag it before it becomes part of a prompt or triggers a tool. It is the gate that stops an attack from ever reaching the model.

Reach for it when

• User or retrieved content flows into your prompts: indirect injection, OWASP LLM01.
• You handle PII and need it flagged before it reaches the model or your logs.
• You want a managed classifier maintained against new attacks, not a regex you own.
• You are provider-agnostic and want one guard in front of any model.

Skip it when

• You have a fully offline, no-external-calls requirement: a hosted API is out.
• You can tolerate open-source classifiers like Llama Guard or Rebuff instead.
• You have zero untrusted input: nothing hostile ever reaches the prompt.

A minimal look

Guard the input, then call the model.

import OpenAI from "openai";
const openai = new OpenAI();

async function answer(userText) {
  // screen untrusted input BEFORE it becomes a prompt
  const verdict = await guard(userText);
  if (verdict.flagged) throw new Error("blocked: prompt injection");

  return openai.chat.completions.create({
    model: "gpt-4o",
    messages: [{ role: "user", content: userText }],
  });
}

The principle it teaches

Where it fits your labs

This is a Week 4 tool. When Lab 4 asks the agent to refuse an injected instruction, Lakera is the input gate that catches the attack rather than trusting the model to notice it.