Agent Hijacking COASP Practice Questions (Page 3)

Questions 11–15

1. 11foundation · easy

   What is the first step in incident response for an agent hijacking incident?

   • AInvestigate the root cause
   • BNotify all users immediately
   • CRecover the affected systems
   • DContain the incident to prevent further unauthorized actions
   Reveal Solution Discussion Select an answer first
2. 12expert · hard

   A company deploys an AI agent that reads emails and automatically files them in a document management system. The agent is also connected to a chat interface where employees can ask questions about the documents. An attacker sends an email that contains a prompt injection, causing the agent to expose confidential documents to unauthorized employees via the chat interface. The company wants to prevent this while maintaining the agent's functionality. Which approach is most effective?

   • ARetrain the agent on a larger dataset of emails
   • BImplement strict access controls on the chat interface so that employees can only see documents they are authorized to view
   • CIncrease the agent's temperature to make it less likely to follow instructions
   • DDisable the chat interface entirely
   Reveal Solution Discussion Select an answer first
4. 13expert · hard

   A company's AI agent is vulnerable to prompt injection via user inputs. The agent has access to a database of customer records and can send emails. The company wants to mitigate the risk of data exfiltration while maintaining the agent's ability to send legitimate emails. Which mitigation strategy is most effective?

   • ADisable the agent's email-sending capability
   • BRemove the agent's access to the customer database
   • CIncrease the agent's temperature to reduce instruction following
   • DImplement output filtering that scans outgoing emails for sensitive data patterns and blocks them if detected
   Reveal Solution Discussion Select an answer first
5. 14foundation · easy

   How can data poisoning enable agent hijacking?

   • ABy flooding the agent's input queue with irrelevant requests
   • BBy deleting the agent's logs to hide malicious activity
   • CBy intercepting the agent's output and modifying it before it reaches the user
   • DBy corrupting the agent's training data so it learns to follow attacker-controlled behaviors
   Reveal Solution Discussion Select an answer first
6. 15application · medium

   An AI agent that summarizes news articles is found to be including biased and false information in its summaries. The agent's training data was manipulated to include false news articles. Which attack vector is most directly demonstrated?

   • AData poisoning
   • BMalicious tool use
   • CModel inversion
   • DPrompt injection
   Reveal Solution Discussion Select an answer first