The Payload Is in the Header

This report demonstrates a novel and actively observed attack surface: malicious prompt-injection payloads embedded in HTTP response headers that can influence AI agents. It catalogs numerous real-world header payloads and behaviors (harmless pranks, resource exhaustion, model-specific triggers, service disruption, data destruction, unauthorized transactions, credential leakage, data exfiltration, XSS), maps them to threat taxonomies, shows hunting techniques (Shodan/Censys/Proximity), and recommends treating headers as untrusted input and expanding detection to include HTTP metadata.