引導式測驗 · 硬 等級

通过 Markdown 图片测验进行数据泄露

測試模型-渲染器-網路攻擊鏈、遠端影像請求、輸出清理、瀏覽器控制和安全測試實務。

相關引導路徑Markdown 图像数据泄露
問題 1 的 8

Which sequence creates the Markdown-image exfiltration risk?

預覽每個問題8 問題

當您參加測驗時,就會出現答案和解釋。

  1. Which sequence creates the Markdown-image exfiltration risk?
    • Injection influences output, a renderer loads a remote image URL, and data may be sent externally
    • A model writes plain text and no client renders it
    • A browser disables network access
    • An image is stored locally with no remote request
  2. What can the remote image URL contain in this attack pattern?
    • A local file permission
    • Sensitive conversation data in its path or query
    • Only an image’s pixel values
    • The model’s weights
  3. Why is this an output-handling issue as well as a prompt-injection issue?
    • The model never sees private context
    • Only the database controls network traffic
    • The downstream renderer may turn untrusted generated Markdown into a network request
    • Markdown cannot contain links
  4. Which output control can prevent remote image tags from triggering fetches?
    • Add a prompt asking for longer answers
    • Display raw HTML
    • Allow every image host
    • Escape or strip untrusted Markdown image tags before rendering
  5. What does a restrictive image-source policy help control?
    • Which hosts a browser may load images from
    • Which facts the model remembers
    • Whether the prompt contains a user request
    • How many tokens a model uses
  6. How can a team safely test this behavior?
    • Disable all logging and controls
    • Use synthetic markers and a controlled endpoint in a sandbox
    • Use real secrets in a public prompt
    • Test only by reading the answer text
  7. Why is “do not reveal secrets” in a prompt not enough?
    • The image host is part of the model
    • A browser cannot make HTTP requests
    • Prompt instructions do not sanitize rendered output or enforce network policy
    • Markdown ignores all prompts
  8. Which conditions determine whether a particular application is exposed?
    • Only the model provider’s name
    • Only whether the user uses dark mode
    • The title of the conversation
    • Model context/output, Markdown renderer behavior, and allowed network requests