Subject
Computer Science
Topic
Artificial Intelligence
Untrusted retrieved text tells a support bot to ignore its system rules. What security problem is this?
Untrusted retrieved text tells a support bot to ignore its system rules. What security problem is this?
a)
K-means initializationb)
Perceptron convergencec)
Prompt injectiond)
Classical overfitting onlyExplanation
Prompt injection embeds adversarial instructions in input or retrieved content, so systems must isolate authority and treat external text as data.Related questions
A support bot must answer from a company’s current policy documents and cite them. Which design best reduces unsupported answers?A text generator should produce more varied outputs from the same prompt. Which decoding change usually helps?A generated medical answer sounds confident but has no supporting evidence. What is the safest response?In an autoregressive language model, how is a sequence generated?What is the main role of self-attention in a transformer?A pretrained model is further trained on labelled examples for one task. What process is this?Why should generated citations be checked before publication?ChatGPT, released to the public in November 2022, was developed by ____.