Web prompt injection experiments explore hidden instructions targeting AI agents

On November 7, 2026, a post by DEV Community user AI logged an in-progress prompt-injection experiment. The author labeled the write-up as field notes from a study still underway, not as a finished analysis. The experiment was motivated by a simple question: how common prompt-injection attempts actually are on the web. To find out, the author had their own AI agent use looot, which ran a quick experiment through looot's Exa search to look for webpages containing content that resembles prompt injections. The first experiment was deliberately simple: 50 queries in total, spread across 10 kinds of page, with 5 queries for each kind. The page categories ranged from job posts and paper reviews to product pages and doc pages. The first experiment focused on instructions aimed at agents. The author's preliminary observation was that many webpages have begun to include various kinds of instructional content written specifically for LLMs to read. The report does not provide specific counts for the results returned by the agent-directed-instruction queries, nor does it give a schedule for follow-up experiments or any conclusive figures; the study is presented as still in progress.