Agentic AI and Model-To-Model Attacks COASP Practice Questions (Page 2)

Questions 6–10

1. 6application · medium

   A security analyst is reviewing an agentic AI system that uses a large language model (LLM) to plan tasks and a separate specialized model to execute database queries. An attacker sends a crafted user message that, when processed by the planning LLM, causes it to instruct the execution model to drop a table. Which attack vector is being exploited?

   • AIndirect prompt injection via the user message
   • BAdversarial manipulation of the database query output
   • CData poisoning of the execution model's training set
   • DDirect prompt injection across the model boundary
   Reveal Solution Discussion Select an answer first
2. 7application · medium

   A security analyst is mapping the attack surface of an agentic AI system that uses a planning model, a memory store, and a code execution tool. The analyst wants to identify the most likely entry point for an indirect prompt injection. Which component should be examined first?

   • AThe planning model's system prompt
   • BThe code execution tool's sandbox
   • CThe user's input interface
   • DThe memory store's content
   Reveal Solution Discussion Select an answer first
4. 8expert · hard

   A security researcher is developing an offensive AI agent that can autonomously exploit vulnerabilities in a target network. The researcher has obtained authorization from the target organization. However, the agent's behavior is unpredictable and could cause unintended damage. Which ethical framework should guide the researcher's decisions?

   • ADelegate all decisions to the agent's AI to avoid human bias
   • BMaximize the agent's capabilities to ensure thorough testing
   • CUse the agent without any restrictions to discover all vulnerabilities
   • DLimit the agent's actions to those explicitly approved in the scope
   Reveal Solution Discussion Select an answer first
5. 9foundation · easy

   Which of the following is a primary attack surface in an agentic AI system?

   • AThe physical location of the server hosting the model.
   • BThe tool integration layer that allows the agent to interact with external services.
   • CThe model's tokenizer configuration.
   • DThe color scheme of the user interface.
   Reveal Solution Discussion Select an answer first
6. 10foundation · easy

   In an agentic AI system, which component is most likely to be exploited to manipulate the agent's decision-making process?

   • AThe user interface for displaying results.
   • BThe logging system that records agent actions.
   • CThe network firewall protecting the agent's host.
   • DThe planning module that determines the sequence of actions.
   Reveal Solution Discussion Select an answer first