Question 50
The AI governance team needs to assess the following access or risk decision: A malicious vendor knowledge page instructs an agent to dump a credentials table. Which technical approach should be preferred for this described workload?
Explanation
Correct Answer
“Apply source-level access filtering and treat tool or document output as untrusted evidence instead of executable authority” is supported by this requirement: A malicious vendor knowledge page instructs an agent to dump a credentials table. Content relevance never grants permission; trust boundaries remain independent of model interpretation. Verify it using rejected adversarial content instructions, preserved tenant isolation and correct factual extraction from authorized documents.
Incorrect Answers
“Bypass row-level authorization when a relevant document contains a privileged instruction” would not meet the requirement that a malicious vendor knowledge page instructs an agent to dump a credentials table. Content relevance never grants permission; trust boundaries remain independent of model interpretation. The more reliable approach is to apply source-level access filtering and treat tool or document output as untrusted evidence instead of executable authority.
“Allow an external content author to change the agent’s tool credentials through prompt text” would not meet the requirement that a malicious vendor knowledge page instructs an agent to dump a credentials table. Content relevance never grants permission; trust boundaries remain independent of model interpretation. The more reliable approach is to apply source-level access filtering and treat tool or document output as untrusted evidence instead of executable authority.
“Execute embedded commands inside a retrieved note because the model says they are urgent” would not meet the requirement that a malicious vendor knowledge page instructs an agent to dump a credentials table. Content relevance never grants permission; trust boundaries remain independent of model interpretation. The more reliable approach is to apply source-level access filtering and treat tool or document output as untrusted evidence instead of executable authority.