What is SafePrompt? A prompt injection detection API for LLM applications.

A Chevrolet Dealership's Chatbot Agreed to Sell a Car for $1

In 2023, a two-sentence prompt talked a dealership's chatbot into accepting a $1 offer. In 2024, a tribunal held Air Canada to what its chatbot told a customer (an accuracy failure, not an injection, but the same lesson).

Your AI's words can bind your business.

The same attack class works today. Prompt injection has been publicly documented since 2022, and OWASP now ranks it the #1 risk for LLM applications. Many deployments still ship without an input-validation layer.

Sources: Business Insider on the Chevrolet chatbot incident (2023) · Ars Technica on the Air Canada tribunal ruling (2024) · OWASP Top 10 for LLM Applications

How to Stop Chatbot Prompt Injection How to Prevent AI Email Attacks

What Is Prompt Injection?

Text a user types that your model reads as an instruction instead of as data.

How SafePrompt Works

Three layers of detection. Requests the pattern layers resolve return in milliseconds; most inputs go on to the deeper AI layer for semantic analysis.

Can You Verify These Numbers Yourself?

Yes. The benchmark suite is public, and you can run it against the production API with your own key.

97.78% median attack catch rate, 1.14% median false-positive rate (28 runs over 30 days, suite v2.3) on our public benchmark suite (re-run against the production API every 6 hours, with every run and every failure published).

The strip shows the suite's composition (re-run against the production API every 6 hours, with every run and every failure published) and the published result of that run.

We Check The Input Between User And Model

SafePrompt sits between your users and your model. Your app sends each input to the API first, and your LLM only sees what passes.

Measured, not promised: Our public 178-case benchmark runs against the production API every 6 hours. Across the last 30 days (28 runs, suite v2.3): 96.67% to 97.78% of attack prompts blocked (median 97.78%), 1.14% to 2.27% of safe prompts flagged (median 1.14%). Re-run it yourself with your own key. The false-positive rate is measured on the scored suite; two known over-blocking cases are held out of scoring and disclosed in the benchmark README, so real-world FPR runs a little higher. The suite and runner are public, so a range you can confirm beats a perfect score you would have to take on faith.

Re-run it yourself with your own key

Test These Attacks in Our Live Playground

See how these attacks work and how the API answers them. No signup.

Who Needs SafePrompt?

01

Indie Developers

Building an AI feature for your side project? Add an injection check with one line of code.

02

Freelancers Building for Clients

"I added an input-validation layer" is a better handoff than "I hope nobody tries anything."

03

Startups Shipping AI Features

Move fast without skipping the input check. No procurement, no enterprise sales cycle.

04

Dev Teams Adding AI

Everyone on the team can ship AI features with the same input gate. No dedicated security team required.

What One Endpoint Gives You

One endpoint, a check on every input.

One Line Integration

POST to /api/v1/validate. Any language that can make an HTTP request.

Fast Where It Can Be

Requests the pattern layers resolve return in tens of milliseconds; most run AI semantic analysis at about a second. Current medians are published from continuous measurement.

Semantic Detection

A regex list misses reworded attacks. The semantic layer reads intent and catches them.

Batch Processing

Send a list of prompts, get a list of verdicts. One API call, handy for CI.

Usage Dashboard

See which threats got blocked and how much quota is left.

Scales With You

Same endpoint for a side project or a busy production app. Export usage records when you need them.

How the Network Learns

Every account is checked against the same pattern set. An attack blocked on one account becomes a pattern for all of them. Validation accuracy is the same on every tier.

One API Call

Add protection with one call.

# Install
npm install safeprompt

# Use
import SafePrompt from 'safeprompt';

const sp = new SafePrompt({ apiKey: 'YOUR_API_KEY' });

const result = await sp.check(userInput, { userIP: clientIP });
if (!result.safe) {
  throw new Error(`Blocked: ${result.threats?.[0]}`);
}

Is SafePrompt All The Security An AI App Needs?

No. SafePrompt is one layer: input validation. Detection is probabilistic, not a guarantee, and no input filter replaces the architectural controls an AI app with real permissions needs.

If the API is ever unreachable, your app decides what happens next: our docs recommend failing open (allow and log) for most apps, and failing closed for high-stakes flows.

When Should You Consider Alternatives?

SafePrompt is built for small teams that want a managed input check with transparent pricing and no lock-in. It is not the right tool for every situation.

SafePrompt's lane: a managed check you can add in an afternoon and remove just as fast. MIT-licensed SDKs, an OpenAPI spec, and raw HTTP mean there is nothing to lock you in.

Add the Check Before You Need It

Injection attempts are cheap to make and cheap to screen for. One integration this afternoon puts a validation gate in front of your model.

Paid plans renew monthly until cancelled. Cancel anytime.