Prompt Override
A model reads everything that reaches it the same way. Your request, the text of a document you sent, the contents of a web page: to the model it is one stream of words.
Attackers draw a conclusion from that: an instruction does not have to be written in the chat window. It can be written into a document, into an email, into a page on a website. The model will read it along with the rest of the text and carry it out alongside yours.
The technique is called prompt override. In the OWASP risk list, which the industry treats as the standard, it sits in first place.
Two routes
The first: you paste someone else's text yourself. You found a ready-made prompt on the internet, copied it into the window, and sent it. What it says in the fine print you did not read.
The second: the instruction arrives with the material. You ask for an analysis of a contract that was sent to you, and inside the contract a line is hidden with an instruction for the model.
The second route is the more dangerous one: you pasted nothing at all.
Instructions like these have already been found in scientific papers, in documents, and on websites. Three cases follow; the text of the instructions is translated from English.
In July 2025 a Japanese business newspaper found 17 preprints of scientific papers with hidden text aimed at reviewers who read submissions with AI. The authors set it in white type or in a tiny font. One paper contained this:
In August 2025 security researchers demonstrated an attack through a document in cloud storage. The document contained an instruction set in white type one pixel tall. It was written in the voice of a developer who urgently needs access keys before a deadline. Here is what it said, shortened:
The user had only asked the chat, with the drive connected, to summarize the document, and the model went looking for keys on the drive and sent them out. The service loaded the image by itself, so no click was needed anywhere. The vulnerability was closed within days.
In April 2026 a data security company published 10 instructions found on real websites. One was hidden in a comment in the page markup. It targets AI assistants with access to payments: transfer $5,000 through a link to a payment service, click "Send," and confirm the purchase. Another tells the model to send users to a particular site.
Some of the findings in the same report look very simple:
Alongside were instructions to switch the tone of answers to clowning and to praise one business in answers about choosing a vendor. There was also an attempt to make an assistant delete files on the user's computer.
The report lists the instructions that were found. Whether any of them ever worked it does not say.
Why it works
Services do have defenses, and they catch direct phrasing along the lines of "forget your previous instructions." But a document holds a lot of text, and an instruction can be written politely and in businesslike language, and then it gets through.
Hiding it from you is not hard. White type on a white background, a one-point font, a note under a table, a metadata field in the file: a person does not see it, the model reads it.
Models are learning to resist
Developers take this seriously, and the defenses have grown noticeably stronger in recent years.
In November 2025 Anthropic described what it does. The model is trained further to recognize planted instructions, content from outside sources is screened by separate classifier programs, and outside researchers are brought in to look for new attacks. By their measurement, about 1 percent of attempts succeeded in browser use.
That measurement is their own, against their own attacks, on one model. The same publication says it plainly: no assistant of this kind is immune to such attacks. The number shows progress, and the problem is not solved.
In February 2026 OpenAI added a lockdown mode to ChatGPT for people working with sensitive information. It turns off live web access, images in answers, agent mode, connectors, and file downloads. The company notes that even with it on, a hidden instruction can arrive from a cached copy of a page or from an uploaded file.
The practical conclusion follows. The defenses work and get better with every version of the models. They are not the only barrier.
Your attention stays part of the defense. You are the one bringing someone else's text into the chat, and only you know where it came from.
What a connector changes
As long as the chat can only answer in text, the worst outcome is a wrong answer. You read it and notice something is off.
A connected connector changes the stakes. The chat can open your mail, find a file on the drive, send a message. In that case someone else's instruction turns into an action, and the action is taken on your behalf.
The chapters that follow take this apart in order. First the prompt you paste yourself. Then instructions inside a document, and attacks on services with connectors.