Question 60
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 validation would best distinguish a working solution from a persuasive but incomplete demonstration?
Explanation
Correct Answer
“Rejected adversarial content instructions, preserved tenant isolation and correct factual extraction from authorized documents” 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
“How frequently external page instructions modify credential scope” 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.
“Restricted records revealed during negative tests” 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.
“Number of unauthorized commands executed successfully” 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.