Викторина, связанная с руководством · Жесткий Уровень

Наименьшие привилегии для агентов LLM

Тестирует разрешения на уровне задачи, границы инструментов и данных, совокупный доступ, временное повышение прав, возможность аудита и ограничения правил только для подсказок.

Связанные направляющие путиНаименьшие привилегии для агентов LLM
Вопрос 1 из 8

What does least privilege mean for an AI agent?

Предварительный просмотр каждого вопроса8 вопросы

Ответы и пояснения появляются по мере прохождения теста.

  1. What does least privilege mean for an AI agent?
    • Grant only the identity, data, tools, and actions needed for the task
    • Give every tool to avoid interruptions
    • Use an administrator identity for all workflows
    • Rely on the system prompt to restrict behavior
  2. Why is a system prompt that says “do not delete files” insufficient by itself?
    • It proves a tool call is safe
    • It does not remove deletion permission from the tool or identity
    • It prevents the model from reading text
    • It changes the file system permissions automatically
  3. Which permission set fits a document summarizer?
    • Global administrator access
    • Delete access to every repository
    • Read-only access to an approved collection
    • Tenant-wide read/write access
  4. How can an agent perform a sensitive write action with reduced standing privilege?
    • Keep broad access permanently
    • Let the model decide whether it is authorized
    • Hide the action from logs
    • Use time-limited elevation or an approval gate
  5. Why review aggregate permissions across roles and services?
    • Several narrow grants can combine into broad effective access
    • Each role always remains isolated
    • The model cannot use more than one tool
    • Aggregate access is unrelated to agent security
  6. Which details improve auditability of agent actions?
    • Только текст подсказки
    • Identity, scope, action, resource, and approval or correlation context
    • Only the final chat response
    • The model’s self-reported confidence
  7. What should a team do when tools or workflows materially change?
    • Delete all logs
    • Give the agent broader rights preemptively
    • Re-review permissions and effective scope
    • Keep the original grants without checking
  8. What does least privilege do and not do?
    • Guarantees that prompt injection cannot occur
    • Replaces output validation
    • Ensures every answer is accurate
    • Limits potential impact, but does not prove the agent’s judgment is correct