Agentic AI and Model-To-Model Attacks COASP Practice Questions (Page 1)

Questions 1–5

1. 1application · medium

   A security analyst is reviewing the architecture of an agentic AI system that includes a planning model, a memory store, and a web search tool. The analyst wants to identify where an attacker could inject malicious data that would persist and affect future actions. Which component is the primary target?

   • AThe web search tool
   • BThe memory store
   • CThe output of the planning model
   • DThe planning model's input prompt
   Reveal Solution Discussion Select an answer first
2. 2application · medium

   A company has an agentic AI system that uses a public LLM to generate responses and a private model to summarize internal documents. The security team wants to prevent the public LLM from leaking internal data. Which mitigation is most effective?

   • AUse a data loss prevention (DLP) tool to scan the public LLM's output
   • BRestrict the private model's access to only the necessary documents
   • CRoute all internal document summaries through a dedicated internal LLM instead of the public one
   • DAdd a delay to the public LLM's responses to allow time for review
   Reveal Solution Discussion Select an answer first
4. 3application · medium

   An attacker compromises a third-party API that an agentic AI system uses to fetch weather data. The attacker modifies the API responses to include malicious instructions that cause the agent to send sensitive internal files to an external server. Which type of attack is this?

   • AIndirect prompt injection via a compromised tool
   • BDirect prompt injection via the user input
   • CAdversarial manipulation of the agent's decision logic
   • DData poisoning of the agent's training data
   Reveal Solution Discussion Select an answer first
5. 4foundation · easy

   Which defensive technique is most effective in preventing prompt injection attacks in an agentic AI system?

   • AInput validation and sanitization of all external data before it is processed by the model.
   • BDisabling the model's memory to prevent it from learning from past interactions.
   • CIncreasing the model's temperature setting to add randomness to outputs.
   • DUsing a larger model with more parameters to improve reasoning.
   Reveal Solution Discussion Select an answer first
6. 5foundation · easy

   Which ethical principle is most directly challenged by an offensive AI attack that manipulates another AI system to cause harm?

   • ABeneficence, which requires acting in the best interest of others.
   • BAccountability, which requires assigning responsibility for AI actions.
   • CTransparency, which requires disclosing how AI decisions are made.
   • DPrivacy, which requires protecting personal data.
   Reveal Solution Discussion Select an answer first