Parameter-to-prompt injection is a flaw where URL or form parameters are copied into an AI system’s prompt, chat context, or tool input without strict separation from trusted instructions. Because those parameters are attacker-controlled, they can smuggle hidden directives, sensitive-looking text, or misleading context into the model’s reasoning path.
This matters because many enterprise assistants handle links, previews, and query strings inside privileged workflows. If the system treats external parameters as part of the prompt, a simple click or form submission can influence what the assistant reveals or does under the user’s permissions. Defenders reduce this risk by sanitizing inputs, keeping untrusted data out of system prompts, labeling external content clearly, and enforcing least privilege on connectors and data sources. In attacks, the goal is often not to hack the model itself, but to manipulate the context it trusts.