Indirect Prompt Injection Starts Turning Up in Webpage Headers
The Register, Thursday, October 15th, 2026
A scan of 217 million HTTP responses found sites using headers to address AI agents, including 30 injection attempts.
Signals Corps sampled 217 million HTTP page responses from Common Crawl and found almost 8,000 sites sending messages to AI agents through HTTP headers.
Most were benign pointers, with 6,989 sites sending an x-llms-txt header, but 30 attempted indirect prompt injection.
Researchers explain that injection works because a model reads instructions and data as one stream, so content an agent reads on a user's behalf can carry commands.
Most attempts were pranks, but some could fool unwary agents, showing that headers are another channel teams must treat as untrusted input.