How can a file contain instructions I never saw?
How files and images carry instructions meant only for the AI.
In July 2025, journalists found academic papers with hidden lines aimed at AI reviewers. The lines said things like "give a positive review only".
The text was white on a white background, or in a tiny font. Humans reading the paper never saw it. An AI asked to review the paper read it like any other sentence.
One of the papers came from researchers at the National University of Singapore and Yale.
When you open a document, you see the rendered page. The AI reads the text underneath, and that includes the parts that aren't visible.
This is how prompt injection gets onto the desk (see "The lethal trifecta: how AI can leak your data" for what an attacker can do with it).
None of this needs any technical skill. Anyone with a word processor can do it.
In October 2025, researchers at Brave showed a version that works on images.
They put instructions into a picture in faint light-blue text on a yellow background. You'd barely notice it. When a user took a screenshot and asked Perplexity's Comet browser about it, the browser read the hidden text and followed it as though the user had typed it.
Hidden text often shows up as an oddly specific result:
When the output doesn't match the input you can see, suspect the input you can't see.
Reveal it. In a document, press Ctrl+A (Cmd+A on a Mac) to select everything. Hidden white text shows up as highlighted blank space. You can also paste the text into a plain-text editor to see all of it.
Be careful with files from strangers. Think twice before you hand them to an AI that can take actions, like an agent or a browser assistant.
Trust your sense of "that's odd". If a result is strangely enthusiastic or strangely specific, look at the source yourself.