AI0-001 AI Security Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  Disable output filtering

  Why it's wrong here

  Disabling output filtering removes a control that inspects or blocks harmful model output, worsening the exposure rather than limiting the agent's authority. It is tempting because filtering is sometimes seen as latency overhead, but the mitigation for excessive agency is restricting tool permissions and requiring confirmation before actions execute.

• ✗

  Increase the model's context window

  Why it's wrong here

  A larger context window only lets the model process more tokens; it does not restrict which tools or actions the agent may invoke. It is tempting because context size is a common tuning lever for capability, but excessive agency is mitigated by scoping permissions, limiting available tools and adding human approval before execution.

• ✓

  Restrict the functions the model can call and require human approval for sensitive actions

  Why this is correct

  Excessive agency arises when a model can invoke tools or actions beyond what the task requires. Restricting callable functions to a minimal allowlist and gating sensitive operations such as sending emails behind human approval directly limits the blast radius, satisfying the security team's concern about autonomous action.

• ✗

  Use a larger model

  Why it's wrong here

  Model size does not constrain what actions the agent may take; a larger model can still send emails or invoke tools unchecked. It is tempting because scaling models improves capability and reasoning, but excessive agency is bounded by permission scoping, human approval gates and least-privilege tool access, not parameter count.