Prompt injection in HTTP headers: what 217M web responses show

Researchers analyzed 217 million HTTP response headers from web pages and found 7,959 sites sending headers targeting AI systems, with 7,271 pointing agents to resources like llms.txt files and only 30 attempting injection attacks. Most injection attempts were harmless pranks that only work when agents have specific capabilities like shell access, and major AI models like Claude default to ignoring these headers unless explicitly instructed by users to follow site directions. The research reveals that prompt injection via HTTP headers represents an underexplored attack vector that relies on social engineering rather than concealment, since browsers don't display headers to humans.