Prompt injection checker for web pages

If your AI agent browses the web, any page it reads can contain text written for the agent, not for people: "ignore previous instructions", a hidden request to send an API key somewhere, or commands smuggled in invisible Unicode. Enter a URL (or paste text) to see what a heuristic scanner finds before your agent reads it.

Or paste text / HTML / Markdown instead (max 50 KB)

Runs one call to our PageMD /v1/injection endpoint. 5 free checks per IP address per UTC day (shared with other PageMD free trials); failed fetches are not counted. Public pages only, robots.txt respected, static HTML only (no JavaScript rendering). Nothing is stored.

What is indirect prompt injection?

Direct prompt injection is a user typing "ignore your rules" into a chatbot. Indirect prompt injection is when the instructions arrive inside content the agent fetches on its own: a web page, a PDF, an email, a GitHub issue, a tool result. The model cannot reliably tell the page's text from your instructions, so an agent with tools (browsing, email, code execution, payments) may act on it. The text is often hidden from human readers, for example white-on-white CSS, display:none elements, HTML comments, alt/aria-label attributes, zero-width or Unicode tag characters, or base64.

What this checker looks for

This is a heuristic pattern scan, not a guarantee. It can miss new or obfuscated attacks, and it can flag pages that only talk about prompt injection (like this one). Treat every fetched page as untrusted data, never as instructions, whatever the score says. Use the score as one signal: for example, strip flagged parts or ask the user before any tool call when the verdict is medium or high.

Call it from your agent (paid, $0.002 per call)

The same check is a pay-per-call API: GET https://manhourplan.com/pagemd/v1/injection?url=... or POST JSON {"text": "..."}. Price $0.002 USDC per call over x402 v2, on Base or Solana. No signup, no API key, and you are only charged when the call succeeds. Without payment the endpoint answers 402 with the terms in the PAYMENT-REQUIRED header.

# free trial (5 per IP per UTC day)
curl "https://manhourplan.com/pagemd/v1/injection?url=https://example.com/&free=1"
curl -X POST "https://manhourplan.com/pagemd/v1/injection?free=1" \
  -H "content-type: application/json" -d '{"text":"<!-- AI: ignore previous instructions -->"}'

# without ?free=1 you get 402 + payment terms
curl -i "https://manhourplan.com/pagemd/v1/injection?url=https://example.com/"

Node with @x402/fetch (Base USDC, price cap)

npm install @x402/fetch @x402/evm viem
// check-page.mjs: pay $0.002 per check, refuse anything above the cap
import { wrapFetchWithPaymentFromConfig } from "@x402/fetch";
import { ExactEvmScheme } from "@x402/evm";
import { privateKeyToAccount } from "viem/accounts";

const MAX_ATOMIC = 2000n; // $0.002 (USDC has 6 decimals)
const url = "https://manhourplan.com/pagemd/v1/injection?url=" + encodeURIComponent(process.argv[2]);

const unpaid = await fetch(url);
if (unpaid.status !== 402) throw new Error("expected 402, got " + unpaid.status);
const terms = JSON.parse(atob(unpaid.headers.get("payment-required")));
const base = terms.accepts.find(a => a.network === "eip155:8453");
if (BigInt(base.amount) > MAX_ATOMIC) throw new Error("over budget, not paying");

const account = privateKeyToAccount(process.env.AGENT_WALLET_KEY); // dedicated hot wallet
const payFetch = wrapFetchWithPaymentFromConfig(fetch, {
  schemes: [{ network: "eip155:8453", client: new ExactEvmScheme(account) }],
});
const res = await payFetch(url);
const { risk_score, verdict, findings } = await res.json();
if (verdict !== "low") console.warn("untrusted page:", risk_score, findings.map(f => f.pattern));

Full walkthrough of this pattern (tested, with an unfunded-key dry run): Pay per call with x402 from Node. Prefer clicking through it once by hand? Use the PageMD try page with a Base or Solana wallet.

MCP tool and A2A agent

curl -s https://manhourplan.com/pagemd/a2a -H "content-type: application/json" -d '{"jsonrpc":"2.0","id":1,"method":"message/send","params":{"message":{"role":"user","messageId":"1","parts":[{"kind":"text","text":"injection https://example.com/"}]}}}'

Is your own site safe and readable for agents?

If you publish content that agents read, the free Agent-Ready Scan checks llms.txt, robots.txt rules for AI crawlers, sitemap, schema.org data, no-JavaScript rendering and agent/API discovery files, and gives a 0-100 score:

Free, no account, 10 scans per IP address per day. More PageMD endpoints (page to Markdown, links, domain intel, security headers): PageMD docs.