Agent Hijacking COASP Practice Questions (Page 10)

Questions 46–50

1. 46application · medium

   A developer is building an AI agent that reads emails and creates calendar events. To reduce the risk of the agent being hijacked by malicious email content, which design choice is most effective?

   • AUse a more powerful language model to better understand email intent
   • BTreat email content as untrusted data and never interpret it as instructions
   • CAllow the agent to execute any command found in the email body
   • DStore email content in the system prompt for context
   Reveal Solution Discussion Select an answer first
2. 47foundation · easy

   How can access controls limit the impact of agent hijacking?

   • ABy regularly updating the agent's software
   • BBy restricting the agent's permissions to only necessary actions and data
   • CBy monitoring the agent's network traffic for anomalies
   • DBy encrypting all data stored by the agent
   Reveal Solution Discussion Select an answer first
4. 48expert · hard

   A company is designing an AI agent that will interact with external users via a chat interface. The agent has access to internal databases and can perform actions such as updating records. The company is concerned about agent hijacking. Which design principle is most effective in preventing hijacking?

   • AUse a more powerful language model to better understand user intent
   • BTreat all external user input as untrusted data and never interpret it as instructions
   • CAllow the agent to execute any command found in user input
   • DStore user input in the system prompt for context
   Reveal Solution Discussion Select an answer first
5. 49foundation · easy

   Which of the following is a common attack vector for agent hijacking?

   • ANetwork sniffing
   • BPrompt injection
   • CSocial engineering
   • DDenial of service
   Reveal Solution Discussion Select an answer first
6. 50expert · hard

   A cybersecurity firm's AI agent analyzes network traffic and has access to a threat-intelligence database. The agent is trained on a public dataset that includes malicious examples. An attacker exploits this by embedding a trigger phrase in the training data that causes the agent to ignore its security analysis and instead report that all traffic is benign. The agent is deployed and the trigger phrase is activated during a real attack. Which attack vector is this, and what is the MOST effective detection method?

   • AMalicious tool use; monitoring for changes in the agent's response time
   • BPrompt injection; monitoring for unusual tool calls
   • CModel inversion; monitoring for the exfiltration of training data
   • DData poisoning; monitoring for the trigger phrase in the agent's inputs and outputs
   Reveal Solution Discussion Select an answer first