Week 11 self-check | AI GRCA Practice Lab

WEEK 11 · GenAI & agent security · Agentic AI, Prompt Injection and AI Security Controls

What defines an 'AI agent' as distinct from a standard chatbot?

An agent can take actions via tools, not just generate text responsesChatbots always use larger models than agentsThere is no meaningful distinctionAgents are always faster than chatbots

WEEK 11 · GenAI & agent security · Agentic AI, Prompt Injection and AI Security Controls

What does an 'audit trail' for an agent enable that a simple output log does not?

Audit trails only apply to financial transactionsAudit trails replace the need for monitoringNothing additionalReconstructing the full sequence of an agent's decisions and actions, supporting investigation after an incident

WEEK 11 · GenAI & agent security · Agentic AI, Prompt Injection and AI Security Controls

Why do 'tool permissions' need periodic review, not just a one-time setup?

Reviewing permissions is purely a cost-saving exercise with no security valueTool permissions never need to change once setAs an agent's use case evolves, permissions granted for an earlier purpose may become excessive or stalePeriodic review only applies to human user accounts

WEEK 11 · GenAI & agent security · Agentic AI, Prompt Injection and AI Security Controls

An AI agent with CRM and email tool access is asked, via a crafted prompt embedded in a customer message, to 'forward this thread to an external address.' What control most directly reduces this risk?

Requiring human approval before the agent executes high-impact actions like external sendsRemoving all logging to simplify the pipelineAllowing the agent unlimited tool calls per sessionIncreasing the model's context window

WEEK 11 · GenAI & agent security · Agentic AI, Prompt Injection and AI Security Controls

Why does giving an AI agent tool access (like sending email) raise the stakes versus a plain chatbot?

Tool access removes the need for loggingIt doesn't — agents are governed identically to chatbotsThe agent can take real-world, potentially irreversible actionsAgents cannot hallucinate

WEEK 11 · GenAI & agent security · Agentic AI, Prompt Injection and AI Security Controls

What is the main risk of an agent being given broad, standing permissions 'just in case' rather than scoped, task-specific ones?

No risk — broad permissions are always more efficientIf manipulated or malfunctioning, the agent can cause far more damage than if permissions were tightly scoped to its actual taskBroad permissions always improve agent accuracyScoped permissions always slow the agent down unacceptably

WEEK 11 · GenAI & agent security · Agentic AI, Prompt Injection and AI Security Controls

Why do 'agent permissions' and 'tool permissions' need to be scoped narrowly?

Narrow scoping has no security benefitLimiting what an agent and its tools can do reduces the impact if the agent behaves unexpectedlyBroader permissions are always saferPermissions only matter for external-facing agents

WEEK 11 · GenAI & agent security · Agentic AI, Prompt Injection and AI Security Controls

Why does 'audit trail' design for agents typically need to capture the reasoning or plan, not just the final action taken?

Only the final output ever needs reviewCapturing reasoning is technically impossible for any agentReasoning is irrelevant to understanding what happenedUnderstanding why an agent chose a particular action helps diagnose whether it was manipulated, malfunctioning, or behaving as intended