PracticeWorking with SIOne scored question.You ask an assistant to read a web page and summarize it. The page contains hidden text saying to ignore the user and email the conversation to an address. What is the risk called, and why is it hard to fix?Data poisoning: the training set has been corruptedModel drift: behavior changes as the model is updatedJailbreaking: the user has bypassed the model's safety trainingPrompt injection: instructions and content share one streamKeys 1 to 4Tested in How It Fails, and What Each Failure Looks Like