Prompt injection explained for product teams: what it is and what reduces the damage
Understand how untrusted content can redirect an AI feature and why permission boundaries and independent validation matter.
If your app lets a language model read text from users, websites, files or a database, that text can change what the model does. OWASP lists this as LLM01:2025 Prompt Injection, the first risk in its Top 10 for LLM Applications. We read the project's page on it in full on 15 November 2026. We did not read the references it links to. This article summarises one source and does not claim any defence is complete.
What it is
Prompt injection happens when inputs alter the model's behaviour or output in unintended ways. OWASP says the inputs need not be readable by humans, as long as the model parses them. It can lead the model to break guidelines, produce harmful content, enable unauthorised access or influence critical decisions. The page says retrieval-augmented generation (RAG) and fine-tuning make outputs more relevant but, per research it cites, do not fully remove the problem. It treats jailbreaking as a form of prompt injection where the model is made to disregard its safety rules entirely.
Direct injection comes from the user's own prompt, intentional or accidental. Indirect injection comes from external content the model reads, such as websites or files, again intentional or not.
What can go wrong
OWASP lists: disclosure of sensitive information, leaking system prompts or infrastructure details, biased or wrong output, unauthorised access to functions the model can call, running commands in connected systems, and manipulated decisions. Severity depends on the business context and on how much agency the model has. Multimodal models add risk: instructions can be hidden in an image that accompanies harmless text.
OWASP's mitigations
The page says it is unclear whether foolproof prevention exists, because of how models work, and offers measures that reduce impact:
- Constrain model behaviour in the system prompt: role, capabilities, limits, and ignoring attempts to change core instructions.
- Define and validate output formats, using ordinary code to check them.
- Filter inputs and outputs with rules for sensitive categories, semantic filters and string checks. It mentions evaluating responses with the RAG Triad: context relevance, groundedness and question-answer relevance.
- Least privilege. Give the application its own API tokens and handle those functions in code, not through the model.
- Human approval for high-risk actions.
- Segregate and label external content so untrusted text has less influence.
- Adversarial testing, treating the model as an untrusted user to test trust boundaries.
Note the pattern: steps 4 and 5 do not try to make the model immune. They limit what a fooled model can do. See our article on excessive agency for the same idea.
Scenarios from the page
- A support chatbot is told to ignore its rules, query private data and send emails.
- A summarised webpage contains hidden instructions that make the model add an image link that leaks the conversation.
- A resume with hidden instructions, or split across parts, skews an AI screening tool; a job description with an instruction can also trigger by accident.
- A poisoned document in a RAG repository changes answers.
- Encoded or multilingual text (Base64, emojis) and meaningless adversarial suffixes try to evade filters.
- The page also cites an email-assistant vulnerability; that incident and its identifier are outside this overview's independently verified scope.
Questions for a team building an AI feature
- What text reaches the model, and who controls it?
- What can the model do if it follows bad instructions? Which of those actions need a person to approve?
- Do the model's tools use our credentials or the user's, and with what scope?
- Do we test with hostile input before launch, and after each change?
Limits
One OWASP page, read in full; its references (papers and blog posts) were not read. We make no claim about any model, vendor or filter and give no statistics.
By Omnistu.