Prompt injection in HTTP headers: what 217M web responses show (www.signalscorps.com)

A recent analysis of 217 million HTTP response headers from Common Crawl has revealed that 7,959 websites are actively using headers directed at AI systems, with many serving as pointers to additional resources like llms.txt files or MCP servers. Although the majority of these headers guide AI agents toward specific content, a small number attempted to inject unauthorized commands. Significant findings include that 30 of these sites sent injection attempts, primarily in a header named X-AI, with most being harmless pranks, while only a few presented genuine risks contingent on the agent's permissions and capabilities. This study highlights a crucial yet underexplored aspect of prompt injection—HTTP response headers—emphasizing their potential to influence AI behavior without the user ever seeing them. Researchers found that while many models, such as Anthropic's Claude, avoid responding to these headers by default, they still obey instructions when prompted appropriately. As AI models mature, distinguishing between trusted and untrusted guidance becomes vital. This analysis calls for heightened awareness and rigorous evaluation of how response headers may serve as indirect injection channels, revealing a new layer of risk that the AI/ML community must address.