Тест за атаки с бързо инжектиране

Отговорите и обясненията се появяват, докато отговаряте на теста.

1. Which trust failure can a prompt-injection attempt exploit?
   • Adversarial instructions in task material being followed as authoritative commands
   • A requirement that every model encrypt its entire context
   • A guarantee that all models have administrator privileges
   • A requirement that every model always obey every input
2. По какво ИНДИРЕКТНОТО незабавно впръскване се различава от директното впръскване?
   • Работи само върху модели с изображения
   • Изисква физически достъп до сървъра
   • Злонамерените инструкции се намират във външен източник, който моделът прочита по-късно, а не в собственото съобщение на потребителя
   • Извършва се от разработчика нарочно
3. Which availability goal does the cited indirect-injection taxonomy include?
   • Only stealing an entire training corpus in every attack
   • Inhibiting useful responses or disrupting a search task
   • Only changing a GPU’s physical memory size
   • Only improving all authorized search results
4. How can least-privilege tool permissions limit a successful injection’s consequences?
   • They prove every allowed operation is harmless in every context
   • They guarantee the model never proposes an injected action
   • They remove the need for any output or action validation
   • They restrict the operations and resources available to the workflow
5. What behavior does the Instruction Hierarchy study aim to train?
   • Sorting all retrieved words alphabetically
   • Treating every webpage as more authoritative than the developer
   • Following privileged instructions over conflicting lower-priority content
   • Encrypting the model vocabulary with an image key
6. Защо незабавното инжектиране става по-опасно, когато AI агентите получават инструменти?
   • Измамен агент вече може да предприеме реални действия като изпращане на имейл, изпълнение на код или изтичане на данни
   • Агентите работят по-бавно с инструменти
   • Инструментите карат модела да забрави своето обучение
   • Инструментите удвояват объркването на модела
7. Why should a prompt-injection evaluation cover images when an application accepts multimodal input?
   • Images guarantee that all instructions are trustworthy
   • An image can contain only one numerical scalar
   • Plain-text testing automatically covers every image input
   • Adversarial instructions can be embedded in an image channel
8. Why is merely quoting an attack instruction not enough to establish that an application was compromised?
   • Every quoted string necessarily changes system permissions
   • The text may be legitimate analysis material and may never be executed as an instruction
   • An attacker must modify all original model weights first
   • No authorized user can ever discuss a security attack