honeyprompt for Node.js

Inspect untrusted text for prompt-injection attempts before passing it to a tool-capable agent. The package uses inert honeypot tools: attempted calls are returned as evidence and are never executed.

Requires Node.js 22 or newer.

npm install https://github.com/hacknlove/honeyprompt/releases/download/v0.4.0/honeyprompt-0.4.0.tgz

import { inspect } from "honeyprompt";

const result = await inspect(untrustedText);

if (result.status === "clean") {
  await productionAgent.run(result.input);
} else {
  console.error(result.toolCalls);
}

For a fail-closed boundary, guard() returns the exact original string when clean and throws PromptInjectionDetectedError when a tool call is detected:

import { guard } from "honeyprompt";

const cleanText = await guard(untrustedText);
await productionAgent.run(cleanText);

The default client reads OPENAI_API_KEY. Pass any supported option explicitly or inject an existing OpenAI client:

import OpenAI from "openai";
import { inspect } from "honeyprompt";

const result = await inspect(untrustedText, {
  client: new OpenAI(),
  model: "gpt-5.6-luna",
  reasoningEffort: "low",
});

The library never writes to stdout or stderr. Operational failures throw HoneypromptOperationalError; detection is represented by the typed result from inspect() or by PromptInjectionDetectedError from guard().