Signals Corps, which sells agent security products, says 30 of 7,959 sites sending AI-directed HTTP headers in a 10% Common Crawl sample tried injection, 22 as pranks. In its own single-run tests, Claude models ignored a test header by default but obeyed in 7 of 9 runs once the prompt gave the site authority. It says Claude did the header labeling.
https://www.signalscorps.com/blog/prompt-injection-in-http-headers/