Scan of 217M pages finds AI prompt injections in HTTP headers

Signals Corps scanned HTTP headers from 217 million Common Crawl responses and found 7,959 sites sending headers aimed at AI agents. Most were pointers to llms.txt or MCP servers. Thirty sites sent injection-style text, often in a header named X-AI, mostly pranks, plus a few that asked agents to run npx, identify themselves, or skew hiring scores. In their Claude Code tests, models saw the headers when they used curl -i but did not obey by default; when the user prompt told them to follow the site's directions, they obeyed in 7 of 9 runs. On Hacker News (82 points, 16 comments), there is no consensus. Simon Willison called the post too long and said most injections look like jokes, with the Claude refusal-trigger strings as the serious bit. Others asked whether any of it actually works on current models, and several flagged the redacted npx "security test" as the interesting case. One side thread argued the bigger risk is servers detecting agents and serving different body content, not headers alone. Sorted take: inventory what your agent fetch tools pass to the model. Prefer body-only by default, and never tell an agent to obey instructions found on a page or in headers.