Guide

What Is Prompt Injection? A Plain-English Guide

The one-sentence answer

Prompt injection is planted text that tricks an AI into doing something you didn't ask for. OpenAI calls it "a type of social engineering attack specific to conversational AI," where a third party, not you and not the AI, slips misleading instructions into what the AI is reading (Source 1).

How prompt injection works (phishing, but for the AI)

Think of a phishing email, except the target is your assistant instead of you. OpenAI makes the same comparison: phishing tries to trick people, and prompt injection tries to trick AIs (Source 1).

Your AI assistant reads lots of things for you: web pages, emails, PDFs, reviews. To the AI, all of that is text, and it can struggle to tell "information to read" from "instructions to follow." Unit 42 puts it plainly: AI models "cannot distinguish instructions from data inside a single context stream" (Source 3).

Hypothetical example: you ask an assistant to summarize your inbox. One email contains a line, invisible to you, telling the assistant to forward your recent messages somewhere. If the assistant treats that line as a command instead of content, it's been injected.

Those planted instructions don't have to be visible. OWASP notes that prompt injections "do not need to be human-visible/readable, as long as the content is parsed by the model" (Source 2).

Direct vs indirect prompt injection

Direct injection is typed straight into the AI. Indirect injection is hidden in something the AI reads for you. - Direct: someone types instructions into a chatbot to make it misbehave, on purpose or by accident (Source 2). - Indirect: the instructions sit in an outside source the AI pulls in, such as a website or a file (Source 2). Our take: this is the one that matters most for everyday users, because you never see the attack. You just asked for a summary.

Prompt injection examples in the real world

Yes, it's happening in the real world. Unit 42, Palo Alto Networks' threat research team, says its telemetry shows indirect prompt injection "is no longer merely theoretical but is being actively weaponized" (Source 3). Among the cases it documented: - A scam ad with hidden text aimed at AI ad-review systems, telling them to approve it. Unit 42 reported this in December 2025 and called it the first real-world case of its kind it knew of. It also says it isn't aware of such an attack actually succeeding against a deployed ad checker (Source 3). - Pages that tried to push agents into payments, such as buying a product, signing up for a paid plan or sending money (Source 3). - Pages that tried to make an AI leak information or run destructive commands (Source 3).

In Unit 42's data, the most common attempts were low-stakes: the top intent, at 28.6%, was simply making the AI produce irrelevant output (Source 3). Our take: many look low-stakes. Some aren't.

What it can lead to

OWASP lists the possible results (Source 2). In plain English: - Leaking your private information to someone else - Revealing the AI system's own hidden settings or instructions - Giving you wrong or slanted answers - Using tools the AI has access to without your OK, like email or payments - Running commands in connected systems - Skewing decisions the AI is helping with

Our take: the more an assistant can do for you, the more an injection can do too.

How to lower your risk

Give specific instructions, limit what the AI can access, and confirm anything sensitive yourself. No tool or setting fully prevents prompt injection; OWASP says it's "unclear if there are fool-proof methods of prevention" (Source 2). But these habits make it harder: 1. Be specific. "Summarize the three emails from my accountant" is safer than "go through my email and handle whatever needs doing." Our take: a narrow request leaves hidden text less room to steer the AI. 2. Limit access. Connect only the accounts a task needs. OWASP recommends least-privilege access (Source 2). 3. Read the confirmation. When an assistant asks before sending, buying or sharing, check that the action and the details match what you wanted (Source 2). 4. Our take: be suspicious of surprises. If an assistant suddenly suggests a link, a payment or a login you didn't ask about, stop and check it yourself. 5. Our take: keep sensitive data out of reach where you can: card numbers, SSN, passwords.

Our take on "is it a scam?": if your AI assistant pushes you toward a payment, a login page or a "verification" form you didn't request, treat it like a phishing message. Don't continue. Go to the company's site yourself. See our phishing training and AI agent shopping safety guides.

FAQ

What is prompt injection in simple terms? Planted text that tricks an AI into doing something its user didn't ask for. OpenAI calls it social engineering aimed at AI (Source 1).

What's the difference between direct and indirect prompt injection? Direct is typed into the AI. Indirect is hidden in content the AI reads, such as web pages, emails or files (Source 2).

Can prompt injection be invisible? Yes. OWASP says it doesn't need to be human-readable if the model processes it (Source 2).

Is prompt injection happening in the real world? Yes. Unit 42 documented websites carrying hidden AI instructions aimed at things like ad-review bypass, payments and data leaks (Source 3).

Is there a setting that stops it? No setting fully prevents it. Specific instructions, limited access and confirming sensitive actions reduce the risk (Sources 1, 2).

Sources (all checked 2026-10-24)

  1. OpenAI, "Understanding prompt injections." https://openai.com/safety/prompt-injections/ — accessed 2026-10-24
  2. OWASP Gen AI Security Project, "LLM01:2025 Prompt Injection." https://genai.owasp.org/llmrisk/llm01-prompt-injection/ — accessed 2026-10-24
  3. Palo Alto Networks Unit 42, "Fooling AI Agents: Web-Based Indirect Prompt Injection Observed in the Wild." https://unit42.paloaltonetworks.com/ai-agent-prompt-injection/ — accessed 2026-10-24 (vendor research; site addresses deliberately omitted here)

OpenAI, OWASP, Palo Alto Networks and Unit 42 are trademarks of their owners, used only to identify them. FraudMeNot is not affiliated with or endorsed by them.

Get the free FraudMeNot newsletter

One weekly scam story and one habit that can help stop it. Free. Unsubscribe anytime.

Sign up for the free newsletter →

Disclaimer. General education only, not legal, financial, tax or security advice. No guarantee. Consult a qualified professional. FraudMeNot content, including the Close Call Report, is general educational and awareness material only. It is not legal, financial, tax, compliance, or cybersecurity advice, and it is not a security service or assessment. Nothing here creates a professional or advisory relationship. Content is provided "as is" without warranties; it may contain errors or become outdated, and no step described can guarantee prevention of fraud or loss. Consult a qualified attorney, accountant, financial professional, or security provider about your specific situation. If you've been scammed, contact your bank immediately and report it at reportfraud.ftc.gov and ic3.gov.

Story labels. Stories are labeled Real, reported, Composite/dramatized, or Hypothetical. People named in reported cases who haven't been convicted are presumed innocent; allegations are only allegations.

Affiliates/sponsors: none in this article.

Questions or corrections: contact us

Labels on this page: Real, reported for the in-the-wild examples (from Palo Alto Networks Unit 42, cited below); Hypothetical for the inbox example, marked where it appears.

Disclosure: This post contains no affiliate links. Some links on FraudMeNot may be affiliate links; if you buy through them, FraudMeNot may earn a commission at no extra cost to you. No product can guarantee you won't be scammed. Details.