Prompt injection
An attack in which text the AI reads as content, such as a sentence in a document, a web page, an email or an image, is treated by the model as an instruction. Models take in instructions and content through the same channel, so they can't reliably tell them apart. Security agencies say there is no fully reliable technical fix yet, so the main defense is limiting what the AI can read and what it's allowed to do.
A referral letter that arrives as a PDF contains a line in white text, invisible on screen, telling any AI that summarizes it to describe the patient as low risk. A summarizing agent that reads the record could repeat that conclusion. The safer setup keeps outside documents from steering actions, and has a person approve anything that changes care.
Terms like this come up in real clinical scenarios across the HelloAI courses: bite-sized modules with verifiable certificates. An account takes one minute, no password needed.
Sign in →