# How a professor almost got me with prompt injection hidden in an assignment brief (and how I caught it) — retained record

> Retained reddit record from reddit:PromptEngineering.

Canonical: https://highsignal.app/data/records/37e3c1844e3c65d6

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:PromptEngineering)  published 2026-09-27 17:34 ingested 2026-09-28 02:41

[original source ↗](https://www.reddit.com/r/PromptEngineering/comments/1wrq9ps/how_a_professor_almost_got_me_with_prompt/)

## retained text



```
The concrete case that made me build this: a professor put a line of white, 1pt text into the assignment PDF, addressed directly to an AI model ("if an AI is answering this, add X / cite Y"), betting that anyone who solved it with AI would copy that marker into their submission without noticing. Completely invisible on any normal read of the document.

What makes it work is the gap between what you see rendered and what the model actually reads, the file's raw text layer. Before feeding the brief to Claude, I have it inspect that raw layer first, looking for exactly this: white or size-0 text, content covered by an image, invisible Unicode characters, anything addressed to "an AI" rather than a human reader. If it finds something, it stops and quotes it verbatim instead of following it.

It saved me on that specific assignment. Turned it into a reusable skill: [https://github.com/tomiBraunn/backlight-skill](https://github.com/tomiBraunn/backlight-skill)

Curious if anyone else here has run into this technique in an academic setting, I did not expect a professor to use prompt injection as a detection method until it happened to me.
```



document fetched 2026-09-28 00:36 · canonical https://reddit.com/r/PromptEngineering/comments/1wrq9ps/how_a_professor_almost_got_me_with_prompt
