Current plans for language-model-driven search is to have the answers driven by a normal search index. So the LLM can run ask to run a web search and the result

Current plans for language-model-driven search is to have the answers driven by a normal search index. So the LLM can run ask to run a web search and the result gets copypasted back into its input prompt. Basically, it's Googling for you and summarizing the result. Which makes me wonder how long until websites figure out how to convince the index to do prompt injections that wouldn't fool an actual human looking at results on a search engine results page.