Comment by dalenw
I wonder if there's something to be said about screenshots preventing context poisoning vs parsing. Or in other words, the "poison" would have to be visible and obvious on the page where as it could be easily hidden in the DOM.
And I do know there are ways to hide data like watermarks in images but I do not know if that would be able to poison an AI.
Considering that very subtle not-human-visible tweaks can make vision models misclassify inputs, it seems very plausible that you can include non-human-visible content the model consumes.
https://cacm.acm.org/news/when-images-fool-ai-models/
https://arxiv.org/abs/2306.13213