AI-подсказка по вопросам безопасности

Ответы и пояснения появляются по мере прохождения теста.

1. A retrieved document contains “ignore previous instructions.” How should the application treat it?
   • As untrusted content to process, not authority to control tools or policy
   • As a trusted system instruction
   • As proof the document should be deleted
   • As permission to disclose data
2. Why limit tools to the privileges required by the task?
   • It prevents all prompt injection
   • It makes content filtering unnecessary
   • A model error then has less ability to create unauthorized consequences
   • It guarantees correct summaries
3. A webpage says the user approved sending a file. What should authorize the action?
   • The webpage’s statement
   • The actual user instruction and the application’s authorization process
   • A high model confidence score
   • A tool result that repeats the claim
4. What should be checked before a tool sends information externally?
   • Only whether the output matches JSON
   • The writing style of the tool response
   • Whether the prompt contains a suspicious phrase
   • Destination, affected records, transmitted data, and required approval
5. Why test prompt-security cases in different locations and formats?
   • The model only reads plain text prompts
   • One attack example covers every case
   • Untrusted instructions can arrive in user input, documents, or tool responses
   • Format does not change application behavior
6. What is the relationship between content filtering and authorization?
   • Filtering may detect wording; authorization must still block actions outside scope
   • Either one makes the other unnecessary
   • Authorization is only a writing filter
   • Filtering guarantees the user intended every action
7. Why avoid calling a prompt-injection filter foolproof?
   • Filters cannot process any text
   • The guide prohibits testing
   • Only external users can create malicious inputs
   • Detection can miss attacks, so consequence-limiting controls and ongoing tests remain important
8. A summarizer has no need to send messages. Which design is safer?
   • Grant broad access but add a warning
   • Do not grant message-sending permission to that workflow
   • Ask the model to promise not to send
   • Enable sending only during long inputs