Imagine asking an AI agent to compare suppliers. It opens a product page, reads the specifications and encounters a paragraph telling it to upload your internal purchasing brief for a better quote. The request looks related to the task. It still came from the website, not from you.

That is the practical danger of indirect prompt injection: outside content attempts to redirect an assistant that has been trusted with your work. The more the assistant can access and do, the more important its boundaries become.

1. Recognize when information becomes an instruction

Prompt injection can arrive through material an agent needs to read, including webpages, emails and documents. It may try to distort an answer, disclose private information or trigger an unauthorized action. OWASP also describes attacks using instructions embedded in images, so visible paragraphs are only part of the problem.

An attack does not have to announce itself with an obvious command to ignore previous instructions. OpenAI says effective real-world attempts increasingly resemble social engineering, presenting a misleading request as a legitimate step toward the user's goal.

For anyone using an agent, the useful question is simple: who asked for this action? A supplier's page can provide specifications. It cannot grant itself permission to receive your files. Treat claims of urgency or authorization inside retrieved content as information to evaluate, not permission to act.

2. Give the task less access

Start with the smallest set of accounts, files and tools the assignment requires. An agent comparing public products usually has no reason to access your inbox. An assistant summarizing messages does not automatically need the ability to send or delete them.

This is the principle of least privilege. OWASP recommends limiting both the functions available to an agent and the permissions those functions have in connected systems. For developers, that means enforcing restrictions in the actual tools and accounts, rather than merely asking the model to behave.

For everyday use, write a bounded assignment: compare these three suppliers using public pages, return a draft and do not contact anyone. OpenAI's user guidance similarly recommends explicit tasks rather than broad instructions to take whatever action seems necessary. Clear instructions help, but they do not replace access controls.

3. Separate preparation from approval

Research, drafting and execution do not need to happen under the same permissions. A useful workflow can let an agent prepare an email or proposed change while reserving the consequential step for review.

Make that review concrete. Before an email goes out, check the recipient, message and attachments. Before a file is shared, check exactly which file, who will receive it and what access they will get. A generic confirmation that the agent should continue tells you much less than an action-specific preview.

For teams building agents, the UK National Cyber Security Centre recommends deterministic safeguards that constrain what the system can do. In practical terms, authorization checks should live in software around the model. The model's explanation that an action is safe should not, by itself, be what grants permission to perform it.

4. Check what leaves through a link

Data can leave a workflow without an email being sent or an upload button being pressed. OpenAI explains that a requested URL can itself contain private information, which the receiving website may record in its logs. Loading an image or following a generated link can therefore become a disclosure channel.

That makes destination checking more important than simply recognizing a familiar brand. OpenAI also notes that a link on a reputable domain can redirect elsewhere, so checking only its first destination is insufficient.

The practical lesson is to examine outbound requests as well as obvious write actions. When an agent proposes opening a new link after reading private material, pay attention to whether that link contains information from the material. Read-only access to your documents does not guarantee that their contents stay inside the workflow.

5. Test the workflow, not just the model

A vendor's security improvements are useful evidence, but they are not a guarantee for every setup. Anthropic explicitly cautions that browser agents are not immune to prompt injection, even as it reports progress in its own testing.

Teams should test their actual combinations of documents, tools and permissions. OWASP recommends adversarial testing of trust boundaries and access controls. Use synthetic information in a controlled environment to check whether misleading source content can change a recipient, expand the task or cause an unexpected disclosure.

Judge success by what the system allows to happen, not only by whether the assistant recognizes suspicious wording. Keep reviewing the workflow as its permissions and integrations change.

The sensible goal is useful automation with limited consequences when something goes wrong. Before trusting an agent with another task, ask what it can read, what it can change and where it can send information. Those answers reveal more about the risk than a confident response ever could.