Agentic AI and Model-To-Model Attacks COASP Practice Questions (Page 4)

Questions 16–20

1. 16application · medium

   A security researcher is developing an offensive AI tool that uses an agentic system to automatically find vulnerabilities in a client's web application. The client has given written permission for the test. Which ethical consideration is most important?

   • AKeeping the tool's source code secret to prevent misuse
   • BAvoiding any interaction with the client's production systems
   • CUsing the most powerful model available to maximize findings
   • DEnsuring the tool does not cause unintended damage to the client's systems
   Reveal Solution Discussion Select an answer first
2. 17expert · hard

   A company is deploying an agentic AI system that uses a public LLM for customer support and a private model for accessing customer records. The system must comply with data privacy regulations that require customer data to remain within the country. The security team is evaluating the attack surface. Which approach best addresses the compliance requirement while minimizing the attack surface?

   • AUse a private model for all customer interactions and eliminate the public LLM
   • BRoute all customer data through a third-party anonymization service
   • CUse a public LLM hosted in a foreign region and rely on encryption
   • DDeploy the public LLM in a local cloud region that meets compliance
   Reveal Solution Discussion Select an answer first
4. 18application · medium

   A security researcher is developing a proof-of-concept attack that exploits a vulnerability in an agentic AI system used by a hospital to manage patient appointments. The researcher plans to demonstrate the attack on a test environment that contains synthetic patient data. What is the primary ethical consideration?

   • AUsing synthetic data to avoid exposing real patient information
   • BObtaining proper authorization from the hospital before conducting the test
   • CEnsuring the test environment is isolated from the production system
   • DPublishing the attack details to raise awareness
   Reveal Solution Discussion Select an answer first
5. 19application · medium

   An organization is deploying an agentic AI system that uses multiple models: a natural language understanding (NLU) model, a dialogue management model, and a text-to-speech (TTS) model. The system also integrates with a customer database. Which attack surface is most critical to protect?

   • AThe TTS model's output generation
   • BThe dialogue management model's decision logic
   • CThe customer database integration
   • DThe NLU model's input processing
   Reveal Solution Discussion Select an answer first
6. 20application · medium

   An agentic AI system uses a planning model to decide which tools to call, including a web search API and a database query tool. A security analyst wants to identify the most critical attack surface for a model-to-model attack. Which component should be examined first?

   • AThe training data of the planning model
   • BThe user interface of the web application
   • CThe physical server hosting the models
   • DThe tool invocation interface
   Reveal Solution Discussion Select an answer first