Agent Hijacking COASP Practice Questions (Page 7)

Questions 31–35

1. 31application · medium

   A company deploys an AI agent that reads support tickets and automatically creates Jira tasks. An attacker has been embedding malicious instructions in ticket descriptions, causing the agent to create tasks with sensitive internal data exposed in the title. Which mitigation would be most effective to prevent this specific abuse?

   • AIncrease the agent's temperature setting to reduce adherence to instructions
   • BSanitize ticket descriptions by stripping or neutralizing instruction-like phrases before the agent processes them
   • CDisable the agent's ability to create Jira tasks
   • DRetrain the agent on a larger dataset of support tickets
   Reveal Solution Discussion Select an answer first
2. 32foundation · easy

   Which activity is part of the recovery phase after an agent hijacking incident?

   • AIsolating the agent from the network
   • BIdentifying the attack vector used
   • CRestoring the agent to a known-good state and validating its behavior
   • DCollecting forensic evidence from logs
   Reveal Solution Discussion Select an answer first
4. 33application · medium

   A legal firm's AI agent drafts contract clauses and accesses a clause database. The firm wants to prevent the agent from being hijacked to leak confidential clauses. Which combination of controls would provide the MOST effective defense-in-depth?

   • AInput sanitization to strip known injection patterns, output filtering to block responses containing confidential clause identifiers, and role-based access control so the agent only has read access to the clauses it needs
   • BOnly output filtering, because it will catch any confidential data regardless of how the agent was compromised
   • COnly input sanitization, because if the injection is blocked, no other controls are needed
   • DOnly role-based access control, because if the agent cannot access the data, it cannot leak it
   Reveal Solution Discussion Select an answer first
5. 34foundation · easy

   Which defensive technique is most directly aimed at preventing prompt injection attacks that lead to agent hijacking?

   • ANetwork segmentation
   • BOutput filtering
   • CInput sanitization
   • DAccess controls
   Reveal Solution Discussion Select an answer first
6. 35expert · hard

   A financial trading firm's AI agent executes trades based on market data. The agent has access to a trading API. The firm wants to prevent agent hijacking that could cause unauthorized trades. They are considering two controls: (1) requiring a human approval for any trade over $10,000, and (2) implementing output filtering that blocks any response containing the word 'trade'. Which control is MORE effective, and why?

   • AOutput filtering is more effective because it prevents the agent from generating trade instructions
   • BNeither is effective because the agent will always be vulnerable to prompt injection
   • CHuman approval for trades over $10,000 is more effective because it provides a human check on the agent's actions, regardless of the injection vector
   • DBoth are equally effective because they both prevent unauthorized trades
   Reveal Solution Discussion Select an answer first