AI-900 Practice Question: Describe Artificial Intelligence workloads and considerations

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  A technique for making AI prompts shorter and more efficient

  Why it's wrong here

  This refers to prompt compression or token optimization, a developer-focused efficiency practice that shortens prompts to save cost or latency. Prompt injection is not about making prompts concise; it is a targeted attack that uses carefully worded instructions to manipulate the model's behavior, irrespective of prompt length. An injected instruction can be extremely short—"ignore previous prompt"—yet be just as dangerous as a long one, so this is unrelated to prompt efficiency.

• ✓

  A security attack where malicious inputs try to override AI system instructions or bypass guardrails

  Why this is correct

  This correctly identifies prompt injection as a security vulnerability. In a prompt injection attack, a malicious user crafts input—often containing instructions like "ignore previous instructions"—to trick the model into bypassing its configured guardrails, accessing prohibited content, or performing unintended actions. Direct injection targets the model's own conversation, while indirect injection can come through third-party data, such as an email or webpage the model later processes. The attack exploits the model's inability to reliably separate data from instructions.

• ✗

  The process of adding examples to prompts to improve model performance

  Why it's wrong here

  This is few-shot prompting or in-context learning, a benign technique where examples are included in a prompt to steer output format or style. Prompt injection is not about providing helpful demonstrations; it is a deliberate attack where input text contains hidden or explicit instructions designed to override the AI's stated guidelines, system prompt, or safety filters. The examples are cooperative, whereas injected content is adversarial and unauthorized.

• ✗

  A method of injecting training data into a model after deployment

  Why it's wrong here

  This describes post-deployment model updating or retraining, which is a legitimate ML lifecycle operation, not an attack. Prompt injection, by contrast, occurs entirely at inference time: an attacker embeds malicious instructions in the text input itself, hoping the model will execute them instead of following its original system prompt. Injecting data into a training set would be data poisoning, a different security threat that alters model weights before deployment.

Go deeper

Related to this question