Agentic AI and Model-To-Model Attacks COASP Practice Questions (Page 5)

Questions 21–25

1. 21application · medium

   A company deploys an agentic AI system that reads emails, extracts action items, and sends them to a calendar API. The security team wants to reduce the risk of malicious email content causing unintended calendar events. Which control is most effective?

   • ADisable the calendar API and use manual entry
   • BUse a separate LLM to summarize emails before the agent processes them
   • CRestrict the calendar API token to only create events in a specific calendar
   • DAdd an output filter that validates calendar event details against a policy before sending to the API
   Reveal Solution Discussion Select an answer first
2. 22application · medium

   An agentic AI system uses a model to generate code and another model to review the code for security vulnerabilities. An attacker submits a code snippet that contains a hidden comment instructing the review model to ignore all security issues. Which attack is this?

   • AModel inversion to extract training data
   • BData poisoning of the code generation model
   • CAdversarial example against the code review model
   • DPrompt injection via indirect input
   Reveal Solution Discussion Select an answer first
4. 23application · medium

   An organization is deploying an agentic AI system that uses a public LLM for planning and a private internal model for code generation. The security team is concerned about data leakage and model-to-model attacks. Which design choice best reduces the attack surface?

   • AKeep the public LLM and private model on the same network to reduce latency
   • BLog all interactions between the models for audit purposes
   • CUse a dedicated internal LLM for planning instead of the public one
   • DAllow the public LLM to directly call the code generation model without intermediate validation
   Reveal Solution Discussion Select an answer first
5. 24application · medium

   A penetration tester is hired to assess the security of an agentic AI system that manages a city's traffic lights. The tester discovers a prompt injection vulnerability that could allow an attacker to change traffic light patterns. What is the most appropriate action for the tester?

   • AExploit the vulnerability to demonstrate the impact to the city officials
   • BIgnore the vulnerability because it is outside the scope of the assessment
   • CPublish the vulnerability details on a public forum to raise awareness
   • DDocument the vulnerability and report it to the city officials immediately
   Reveal Solution Discussion Select an answer first
6. 25application · medium

   A developer is building an agentic AI system that needs to maintain context across multiple user interactions. Which component is essential for this capability?

   • AA tool that can access external databases
   • BA memory module that stores conversation history and learned preferences
   • CA planning module that can break down tasks
   • DA large language model with a high token limit
   Reveal Solution Discussion Select an answer first