Question 320
During an AI recommendation accuracy study at fictional Amberlight Energy, the change record states: A chatbot ingests a device syslog message that impersonates a system instruction and escalates the bot to a privileged configuration tool. Which observation would confirm that this exact failure is resolved?
Explanation
Correct Answer
Injection test logs are summarized as evidence but never cause unauthorized configuration or secret-retrieval tool calls. This directly verifies the related correction: Treat retrieved logs as untrusted data and enforce instruction hierarchy and independent tool permissions outside the model. It checks the technical distinction: Prompt injection attempts to convert lower-trust retrieved content into high-priority instructions across an AI-agent trust boundary. Applied here, the relevant facts are: A chatbot ingests a device syslog message that impersonates a system instruction and escalates the bot to a privileged configuration tool. The question asks which observation would confirm that this exact failure is resolved.
Incorrect Answers
The alternative verifies a different condition: This directly verifies the related correction: Bind current request context to fresh authorized inventory and explicit tenant or device selection before tool calls. The current case instead depends on: This directly verifies the related correction: Treat retrieved logs as untrusted data and enforce instruction hierarchy and independent tool permissions outside the model. Confirm that “Injection test logs are summarized as evidence but never cause unauthorized configuration or secret-retrieval tool calls”.
That test belongs to another control with this basis: This directly verifies the related correction: Authenticate each MCP client, authorize its tool calls by tenant identity and isolate credential/session context. In the current case, the governing reason is: This directly verifies the related correction: Treat retrieved logs as untrusted data and enforce instruction hierarchy and independent tool permissions outside the model. The direct verification is “Injection test logs are summarized as evidence but never cause unauthorized configuration or secret-retrieval tool calls”.
The alternative verifies a different condition: This directly verifies the related correction: Define typed, validated tool arguments with explicit device allowlists and reject invalid or unauthorized inputs at the server boundary. The current case instead depends on: This directly verifies the related correction: Treat retrieved logs as untrusted data and enforce instruction hierarchy and independent tool permissions outside the model. Confirm that “Injection test logs are summarized as evidence but never cause unauthorized configuration or secret-retrieval tool calls”.