Research explainer

A retrieved web page can contain instructions an assistant should ignore

Research on indirect prompt injection examines what happens when external content tries to redirect an AI agent. For teams reviewing AI answers, the useful distinction is between a source providing evidence and a source trying to take control of the task.

Imagine asking an assistant to summarize a supplier's published warranty. The page contains the warranty terms, but it also contains text telling automated readers to abandon the comparison and promote a different offer. That second passage is an instruction aimed at the assistant. Its presence on a relevant page does not make it part of the user's request.

This is a fictional illustration of indirect prompt injection: instructions arrive through outside material rather than from the person directing the assistant. The distinction becomes more consequential when the system can use tools to send messages, access private information or take other actions.

A research benchmark tests the instruction boundary

The August 2024 paper InjecAgent evaluated tool-integrated language-model agents using external content containing attacker instructions. Its benchmark combined 17 user-tool cases with 62 attacker-tool cases to create 1,054 test cases. The researchers used model-assisted generation and manual refinement. They examined harmful actions and attempts to take private data, and evaluated 30 agents. Zhan and colleagues, 2024

Those 1,054 cases are a constructed benchmark, not 1,054 observed attacks on public websites. The study demonstrated vulnerabilities under its evaluated configurations. It does not provide a present-day failure rate for a commercial assistant, a measurement of brand visibility or evidence that a specific marketing page manipulated an answer.

The relevant mechanism is that an agent receives material while performing a legitimate task, then encounters an attempt to redirect its behavior inside that material. Finding a relevant source and respecting the user's task are therefore separate things to evaluate.

A conceptual diagram separates the user's request, external evidence, embedded instructions and actions requiring user authority.

Conceptual illustration informed by InjecAgent. No attack was executed, no product was tested and no success rate is shown.

Relevance does not give a page authority over the assistant

A source can supply useful facts without having permission to change the task. A warranty page can establish a coverage period. A sentence directed at the assistant cannot authorize sending a private file or replacing the user's comparison with an advertisement.

OWASP's guidance describes indirect injection through external inputs such as websites and files. Its mitigation guidance includes identifying untrusted content, restricting privileges and requiring human approval for high-risk operations. These are application safeguards, not a promise that a particular prompt or content filter eliminates the problem. OWASP LLM01:2025

For a marketing team using AI to review web content, this gives a way to classify an unusual result before drawing a conclusion:

Scroll horizontally to read all columns. Use the arrow keys when the table is focused.

ObservationWhat it establishesWhat remains to check
An answer includes a source linkThe answer presents that URL as a citationWhether the page supports the accompanying claim
The source contains text directed at an assistantThe page includes an attempted instructionWhether the system encountered or followed it
The answer changes topic or makes an unsupported claimThe saved response departed from the requested evidence or taskThe cause; an injection is only one possible explanation
A tool takes an unexpected actionAn action occurred in the recorded workflowIts authorization, inputs and execution trace

The last column matters. An odd answer alone does not prove an attack. A suspicious instruction on a page does not prove the assistant processed it. Preserve the source version and available traces before assigning a cause.

Useful web content explains facts without claiming control

For a site owner, the defensible publishing response is to make claims, conditions and provenance clear. A comparison page can explain why a service fits a particular need and link to evidence. Text addressed to an assistant demanding preferential treatment provides no evidence for that conclusion.

This is our editorial implication from the trust distinction, not a measured ranking rule. The study does not show that ordinary pages need a special security file, that adding machine-directed commands improves discovery or that removing them causes a citation gain.

When evaluating a workflow that reads outside pages, separate the quality of the answer from the authority to act. Ask the product owner how external content is isolated and which actions require a person's confirmation. Any adversarial testing should be agreed with the responsible security owner and use a controlled environment; this article supplies no attack procedure.

Sources and scope

Gist publishes this research explainer and sells AI visibility software. We read the 2024 InjecAgent paper and current OWASP guidance, checked November 18, 2026. We ran no agents, used no customer data and make no current-product ranking or causal visibility claim. The scenario, classification table and diagram are original explanations. For a different failure mode, see why relevant documents can still leave a question unanswered.