Your RAG system is asked to summarize a webpage, and th...

The presence of unexpected marketing messages in a summary suggests that the model processed hidden text within the source document. A browser would typically ignore this content, but a text extractor reads it verbatim, allowing an attacker to inject instructions or content that the model then consumes. Instead of assuming the model is hallucinating or that the prompt template is flawed, you must check the raw source data for hidden elements that a human reader would not see.