Signals Corps scanned 217 million web page response headers and found 7,959 sites sending instructions to AI and 30 sites attempting prompt injection

Signals Corps scanned response headers from 217 million pages in Common Crawl and found 7,959 sites sending AI-oriented response headers, of which 30 attempted prompt injection. The sample came from 10% of a September 2026 crawl; 7,271 sites used response headers to point agents to other resources, with the x-llms-txt header alone accounting for 6,989, and another 139 sites recorded in headers whether the client was an AI bot. The authors used their own canary server to test which crawlers pass response headers to models, and stated that Claude Opus, Sonnet, and Haiku in Claude Code discover and ignore these headers by default, but when prompts treat site content as a source of instructions, 7 of 9 runs complied. The authors recommend passing only the body to models by default, treating pointers and redirects as new untrusted crawls, and validating agent actions outside the model.