Half the Validated Prompt Injections in a Web-Scale Corpus Rode in HTTP Response Headers, Where a DOM-Only Scanner Can’t Reach
An indicator-driven study built a corpus from 1.2 billion Common Crawl URLs plus indicator-matched Censys and Shodan snapshots, and 51.2% of its 15,387 validated injections sat in custom HTTP response headers rather than page content. Here is what that number does and does not measure, why header presence is a serialization problem rather than a protocol problem, and why the representation your pipeline hands the model matters more than the header namespace.