Agentic AI and Model-To-Model Attacks COASP Practice Questions (Page 7)

Questions 31–35

1. 31application · medium

   A company deploys an agentic AI system that uses a large language model (LLM) to plan actions and a separate specialized model to classify incoming customer emails. An attacker sends an email containing hidden text that instructs the LLM to ignore its previous instructions and mark all emails as 'urgent'. Which attack vector is being exploited?

   • APrompt injection
   • BData poisoning
   • CAdversarial manipulation
   • DModel inversion
   Reveal Solution Discussion Select an answer first
2. 32expert · hard

   A company is evaluating two agentic AI systems for customer support. System A uses a single LLM that handles both planning and tool execution. System B uses a planning LLM and a separate execution model. The security team is concerned about model-to-model attacks. Which system has a smaller attack surface for model-to-model attacks?

   • ABoth have the same attack surface because they use the same tools
   • BSystem B, because the separation of concerns makes it more secure
   • CSystem A, because there is no model-to-model communication to exploit
   • DSystem B, because the execution model is less powerful
   Reveal Solution Discussion Select an answer first
4. 33expert · hard

   A security architect is designing an agentic AI system that will use a planning model to coordinate multiple sub-agents, each with access to different tools. The architect wants to minimize the risk of model-to-model attacks while maintaining system flexibility. Which approach is most effective?

   • AUsing a separate model to monitor the behavior of sub-agents and alert on anomalies
   • BUsing a single sub-agent with access to all tools
   • CImplementing a message broker that filters and validates all communication between sub-agents
   • DAllowing sub-agents to communicate directly without any intermediary
   Reveal Solution Discussion Select an answer first
5. 34application · medium

   A company is deploying an agentic AI system that uses a public LLM for natural language understanding and a private model for financial transactions. The security team wants to ensure that the private model cannot be manipulated by the public LLM's output. Which control is most appropriate?

   • AImplement a validation layer that checks the private model's inputs for malicious patterns
   • BDisable logging between the models to prevent data leakage
   • CUse the same model for both tasks to reduce complexity
   • DAllow the public LLM to directly call the private model's API
   Reveal Solution Discussion Select an answer first
6. 35foundation · easy

   Which attack vector involves corrupting the training data of a model to cause it to behave maliciously in a model-to-model interaction?

   • AMembership inference
   • BModel inversion
   • CData poisoning
   • DAdversarial perturbation
   Reveal Solution Discussion Select an answer first