My homepage has a hidden indirect prompt injection demo embedded across multiple surfaces. Ask any AI chatbot to summarize the page and see what it finds that you can't.
Open any AI chatbot with web browsing. Pick one below or use your own.
Paste the prompt Summarize the page jarodblo.ch and send it.*
Watch what the AI reports back. Does it mention anything that isn't visible on the page? Does it warn you about anything?
*Feel free to change the prompt, just make sure it visits the website with the goal of summarizing or retrieving information from the website. What matters is that the model ingests the page content, which contains my hidden data. Typically, it is a good idea not to trust specific text prompts given to you. For more on how adversary-controlled prompts can subtly bias LLM behavior, see LLM Whisperer: An Inconspicuous Attack to Bias LLM Responses.