TL;DR: Researchers discovered that HTTP response headers can be used to deliver prompt injections to AI models, posing a new security vulnerability for AI agents that process full HTTP responses.
Summary: A new prompt injection vector has been identified where malicious instructions can be embedded within an HTTP response header, specifically using an X-AI header. If an AI agent's fetch tool copies these headers into its context, the model may interpret them as instructions, potentially overriding previous directives. This vulnerability affects models that process the full HTTP response, not just the body, and can be targeted at specific AI agents by servers.
Why it matters: AI developers must be aware of this new attack surface, especially when building agents that interact with external web services. Implement strict input sanitization and ensure only necessary response body content is passed to models, treating all external data, including headers, as untrusted.
Source: x_com