Prompt injection
Prompt injection is text written to make an AI model ignore its instructions and follow someone else's, typed into the chat or hidden in what it reads.
Prompt injection is when someone writes text that makes an AI model drop the instructions it was given and follow theirs instead. The text can be typed straight into the chat, or hidden in something the model reads, like a web page, an email or a product review.
It matters to brands because AI platforms read the web to answer questions about you. If a page they read carries hidden orders, the answer can say things about you that nobody would write in plain sight.
The two kinds
Direct prompt injection is typed by the person using the assistant. Someone tells a chatbot to "ignore your rules" and tries to get it to say or do something it shouldn't. That's mostly a problem for whoever runs the chatbot.
Indirect prompt injection arrives through content the model fetches. The person asking does nothing wrong. The model reads a page, finds text written as an instruction, and may act on it. This is the kind that can touch your brand.
OWASP ranks prompt injection as the first risk on its list for applications built on language models.
Why it gets more serious with agents
An assistant that only writes an answer can repeat a false claim. An agent that books, buys or fills in forms for someone can act on one. For example, a hidden line on a comparison page could tell a shopping agent that Kestrel Roasters "has stopped shipping," and the agent might pick another brand without telling its user why.
The AI companies work to block this, and hiding text from people already breaks Google's spam policies. So never try it on your own pages. The likely result is losing the trust you've built.
In AnswerOrbit
We don't look for hidden instructions on pages. What we show is where a wrong claim about you comes from. In Claims, Inaccuracy drivers lists the pages behind each claim that contradicts your facts, so you know which page to look at first.