Agent Hijacking COASP Practice Questions (Page 11)

Questions 51–55

1. 51application · medium

   An organization's AI agent that manages customer orders is hijacked via a prompt injection attack. The agent incorrectly marks a large number of orders as 'cancelled' and sends confirmation emails to customers. Which of the following is the most significant consequence of this incident?

   • AIncreased network bandwidth usage
   • BReduced model accuracy on future tasks
   • CPermanent deletion of the agent's training data
   • DLoss of customer trust and potential legal liability
   Reveal Solution Discussion Select an answer first
2. 52expert · hard

   A university's AI agent assists students with course registration and has access to the student records system. An attacker sends a message through the student forum that the agent monitors. The message contains a hidden instruction that causes the agent to change a student's grade to an 'A'. The agent's normal grade-verification logic is overridden. The security team is deciding between two detection methods: (1) monitoring for changes to grades that are not initiated by authorized staff, and (2) monitoring for the specific hidden instruction phrase in the agent's inputs. Which detection method is MORE effective, and why?

   • AMonitoring for the hidden instruction phrase is more effective because it detects the attack at the input stage
   • BBoth are equally effective because they both detect the attack
   • CMonitoring for unauthorized grade changes is more effective because it detects the impact of the attack, regardless of the injection vector
   • DNeither is effective because the agent will always be vulnerable to prompt injection
   Reveal Solution Discussion Select an answer first
4. 53application · medium

   An AI agent that generates code snippets for developers is being hijacked by prompts that cause it to include malicious code in its output. Which output filtering technique would be most effective?

   • AIncrease the model's max token limit to allow more complete code
   • BScan the generated code for known malicious patterns and block the output if detected
   • CDisable the agent's ability to generate code
   • DUse a different prompt template for the agent
   Reveal Solution Discussion Select an answer first
5. 54expert · hard

   An AI agent that manages a hospital's patient appointment system is hijacked. The agent cancels thousands of appointments and sends incorrect confirmation messages to patients. The hospital is now facing a lawsuit from patients who missed critical treatments. Which impact of agent hijacking is most directly demonstrated?

   • ALoss of user trust and legal liability
   • BIncreased network bandwidth usage
   • CPermanent deletion of patient records
   • DReduced model accuracy
   Reveal Solution Discussion Select an answer first
6. 55application · medium

   An AI agent that manages a company's calendar is suspected of being hijacked. Which monitoring approach would best detect unauthorized calendar changes?

   • ALog all calendar changes and alert on any change that is not initiated by a known user
   • BMonitor the agent's CPU usage
   • CCheck the agent's model version daily
   • DReview the agent's training data for anomalies
   Reveal Solution Discussion Select an answer first