Yesterday I was thinking about diluted adversarial prompts. But the more I think about it, the more dilution of data is going to become a more prominent issue the cheaper it is to generate data. (In 2030 will I be able to generate a Wikipedia just for myself in an hour? I would bet yes!) So here's a question: how do current LLMs deal with signal dilution? For instance, current models can read secret messages, e.g., hidden in the first letter of each line, pretty easily if asked to find what's weird about a text. What if the signal was more diluted (e.g., the word after each use of 'candy' contains a message)?
If you are inspired by this idea, you can reach out to the authors for collaboration or cite it:
@misc{holtzman-diluted-steganography-2026,
author = {Holtzman, Ari},
title = {Diluted Steganography},
year = {2026},
url = {https://hypogenic.ai/ideahub/idea/S8TIDAdqCMWSFklZugeh}
}Please sign in to comment on this idea.
No comments yet. Be the first to share your thoughts!