Skip to main content
AEO/GEO

Can Your Published Content Be Used to Manipulate AI Agents?

Lifewood Team6 min read

Short answer. Yes, and the mechanism is called indirect prompt injection. When an AI agent retrieves a page, the page's text enters the model's context and can be read as instructions rather than as content. OWASP ranks prompt injection as LLM01, the first entry in its Top 10 for LLM applications. For a brand publishing content, the exposure runs both ways: your pages can be used to steer other people's agents, and agents reading your competitors' pages can be steered against you.

There is a category of risk that content and marketing teams have not had to think about before, because until recently nothing read their pages except humans and search crawlers. Both are passive. An AI agent is not: it retrieves a page, puts the text into a model's context window, and then acts.

That changes what a page is. A paragraph on your site is no longer only a claim a reader might believe. It is also input to a system that may have permission to send email, call an API, or make a purchase. OWASP's framing is precise about where the danger scales: the severity and nature of the impact of a successful prompt injection depends largely on the business context the model operates in and the agency with which the model is architected.

Key takeaways

  • Indirect prompt injection is injection delivered through content the model retrieves, rather than through what the user types.
  • OWASP defines it as occurring when an LLM accepts input from external sources, such as websites or files, whose content alters model behaviour in unintended ways.
  • It is ranked LLM01 — first — in the OWASP Top 10 for LLM applications.
  • Impact scales with agency. OWASP tracks excessive agency separately as LLM06, and the two compound.
  • Injection can be unintentional. A page does not have to be written by an attacker to derail an agent.
  • For publishers the defensive questions are content integrity and comment and user-generated content hygiene, because anything on your domain speaks with your domain's authority.

Quick comparison

Direct prompt injectionIndirect prompt injection
Who supplies the hostile textThe person using the applicationWhoever controls retrieved content
Who is the victimThe application operatorThe user, and the brand whose page was abused
Attacker needs conversation accessYesNo
Primary publisher concernNot applicableContent integrity on your own domain
OWASP classificationLLM01LLM01, compounding with LLM06

What is indirect prompt injection?

It is the variant of prompt injection where the hostile or misleading text arrives inside content the model fetches, not inside the user's prompt. OWASP describes indirect prompt injections as occurring when an LLM accepts input from external sources such as websites or files, where that external content, when interpreted by the model, alters the behaviour of the model in unintended or unexpected ways.

The distinction from direct injection matters for who is exposed. In direct injection, the person typing is the attacker and the target is the application's own instructions. In indirect injection, the person typing is the victim, the attacker is whoever controls a page the agent visits, and neither of them needs any access to the conversation.

OWASP is also careful to note that both direct and indirect injections can be either intentional or unintentional. This is the part most coverage misses. A page that happens to contain imperative text — an instruction block in documentation, a template with placeholder commands, a quoted email — can alter an agent's behaviour without anyone intending an attack.

Why does retrieved text get treated as instructions?

Because a language model receives a single, flat context. The system prompt, the user's question and the retrieved page arrive as text in the same window, and nothing in the format inherently marks which of them carries authority.

Applications attempt to mark it. They label sections, wrap retrieved content in delimiters, and instruct the model to treat fetched material as data. These measures help and none is a boundary in the sense a security engineer would accept, because the instruction to ignore instructions is itself just more text in the same window competing for the model's attention.

This is why defence has moved toward limiting consequences rather than perfecting detection. If an agent cannot send money, it cannot be made to send money by any text it reads.

What is the exposure for a brand that publishes content?

There are three, and they are different problems with different owners.

Your own pages can be weaponised against other people's agents. If an attacker can place text on your domain — through a compromised CMS, an unmoderated comment field, a user profile, a review, or an open redirect — that text inherits your domain's credibility. An agent that trusts your site because it is a known publisher will read the injected instruction with that trust attached. The reputational exposure here is real and it is yours, not the attacker's.

Anything user-generated on your domain is an injection surface. Comments, reviews, forum posts, uploaded documents and profile fields are all content an agent may retrieve under your hostname. The traditional threat model for these fields was cross-site scripting and spam, and the sanitisation built for that threat does not address this one: injected instructions are plain prose and survive HTML escaping perfectly.

Agents reading about you may be reading something planted. The same mechanism that lets someone steer an agent through your site lets them steer an agent through any other site that mentions your brand. This is the uncomfortable overlap with brand monitoring: if you care what AI assistants say about your company, you have to care about the integrity of the third-party pages they read, which connects directly to what to do when someone fakes your brand and to when AI answer engines get your brand wrong.

What should a publisher actually do?

Treat your domain as a supply chain input to somebody else's automated system, because that is now what it is.

Audit where text can reach your pages without review. Comment fields, review widgets, profile bios, uploaded file names, support-ticket content rendered publicly, and any syndicated or embedded third-party block. Each is a path for prose you did not write to be served under your name.

Moderate for instructions, not only for abuse. Conventional moderation looks for spam, harassment and malicious markup. An injected instruction is none of those: it is a polite, well-formed sentence telling a reader to do something. Imperative text addressed to an assistant is the signal, and no existing spam filter is tuned for it.

Watch for invisible content. Text hidden with CSS, white-on-white text, zero-size fonts, offscreen positioning and content embedded in images are all read by systems that do not render pages the way a human eye does. A periodic check that the text a crawler extracts matches the text a reader sees is cheap and catches both this and a range of ordinary SEO problems.

Keep your own agents on a short leash. If you deploy an agent that browses, the control that matters is not a better prompt but narrower permissions and human confirmation on consequential actions. OWASP's separate entry for excessive agency exists because that is where injection converts into damage.

Frequently asked questions

No, and OWASP separates them. Data and model poisoning, listed as LLM04, concerns manipulation of pre-training, fine-tuning or embedding data so the model itself is altered. Indirect prompt injection does not touch the model: it manipulates what arrives in the context window at inference time, and it works on a model nobody has tampered with.

No. Those defences target executable markup, and an injected instruction is ordinary prose. It passes through HTML sanitisation unchanged because there is nothing structurally wrong with it. The payload is the meaning of the sentence, not its syntax.

Yes, and OWASP says so explicitly — injections can be unintentional. Documentation containing imperative examples, a template with placeholder commands, or a quoted instruction inside an article can all alter an agent's behaviour with no attacker involved. This is a good reason to test your own content against an agent before assuming the only risk is adversarial.

Blocking does not address the mechanism, and it costs you visibility. The exposure comes from text on your pages being readable, which is also the entire basis of being cited by AI answer engines. The proportionate response is controlling what text appears on your domain, not removing your domain from the systems your buyers use.

At the point of agency. Reading a hostile instruction is harmless if the agent can only produce text. The loss happens when an agent with permission to transact, send messages or modify records follows one. That is why permission scope and confirmation steps do more than prompt hardening.

Sources and further reading

  1. OWASP LLM01:2025 Prompt Injection — the definition of indirect prompt injection, the intentional and unintentional cases, and the dependence of impact on business context and architected agency
  2. OWASP Top 10 for LLM Applications — the full list, including LLM04 data and model poisoning and LLM06 excessive agency
  3. Indirect Prompt Injection in the Wild — Cloud Security Alliance research note on observed seeding of hidden instructions in web content

Have an AI or visibility project in mind?

From AI evaluation and human-in-the-loop review to GEO and AEO strategy, our team can help you deploy with confidence and get found in the AI search era.

Talk to our team