A new vector for AI attack is ermging: “drowning”.
Essentially, in a world of data lakes and other large repositories of data, an AI tool may find itself overwhelmed with the volume of data it tries to utilize in order to generate a response. While there are safeguards to force an AI tool to stop and simply prepare what it’s found so far, an engineered prompt can force the AI to keep looking for more, and more, and more.
“It acts like it’s the end of the world if it can’t find the fullest, most definitive answer,” says AI researcher Julie McVoy. “We call it the U2 effect.”
New protections against AI drowning are being developed by leaders in agentic AI management such as Microsoft. An announcement is expected later today (April 1, 2026).