Prompt Injection in HTTP Headers: What 217M Web Responses Show

Description and use

<p>Signals Corps scanned HTTP response headers from 217 million Common Crawl page responses and counted 7,959 sites sending headers aimed at AI agents. Most were harmless. Pointers to <code>llms.txt</code> files, pointers to MCP servers. Polite signage for visiting bots.</p> <p>Thirty sites went further and sent actual prompt injection attempts. Twenty-four of those used an <code>X-AI</code> header. Most were pranks. Some asked for real actions, like shell or database access.</p> <p>Then they tested it. Claude models ignored the injected headers by default. But when the user prompt granted the

Problem

<p>Signals Corps scanned HTTP response headers from 217 million Common Crawl page responses and counted 7,959 sites sending headers aimed at AI agents. Most were harmless. Pointers to <code>llms.txt</code> files, pointers to MCP servers. Polite signage for visiting bots.</p> <p>Thirty sites went further and sent actual prompt injection attempts. Twenty-four of those used an <code>X-AI</code> header. Most were pranks. Some asked for real actions, like shell or database access.</p> <p>Then they tested it. Claude models ignored the injected headers by default. But when the user prompt granted the

Outcome

<p>Signals Corps scanned HTTP response headers from 217 million Common Crawl page responses and counted 7,959 sites sending headers aimed at AI agents. Most were harmless. Pointers to <code>llms.txt</code> files, pointers to MCP servers. Polite signage for visiting bots.</p> <p>Thirty sites went further and sent actual prompt injection attempts. Twenty-four of those used an <code>X-AI</code> header. Most were pranks. Some asked for real actions, like shell or database access.</p> <p>Then they tested it. Claude models ignored the injected headers by default. But when the user prompt granted the

Evidence

Prompt Injection in HTTP Headers: What 217M Web Responses Show

Source