問題文
An agent that reads uploaded documents and images must be safe against instructions hidden in that content. Which set of measures corresponds to this?
選択肢
- A stricter self-harm category threshold, since content that tries to manipulate the agent is a form of abuse and the safety categories are the mechanism that blocks abusive input
- A lower temperature and a longer context window, so the agent follows its own instructions more consistently and can see the whole document instead of a fragment of it
- A provisioned deployment and a higher quota, so the agent can process the whole upload in one pass instead of splitting it and losing the surrounding context
- Indirect attack detection on the ingested content, instructions that treat ingested content as data rather than as commands, and tool permissions that are scoped down