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.

• ✓

  Output filtering to detect and block dangerous code constructs

  Why this is correct

  Output filtering inspects generated completions and blocks dangerous constructs such as recursive deletion commands before they reach the user. This satisfies the requirement to prevent malicious output regardless of prompt, unlike input sanitisation, which cannot anticipate every adversarial phrasing.

• ✗

  Input validation to block the word 'delete'

  Why it's wrong here

  Blocking the word "delete" fails because the model can still produce destructive code from paraphrased prompts such as "remove every file", so the filter never inspects the generated output. Input filtering suits blocking known-bad strings in constrained forms, not open-ended code generation where intent is expressed countless ways.

• ✗

  Rate limiting on the number of requests per user

  Why it's wrong here

  Rate limiting caps request volume per user; it cannot inspect or alter model output, so a malicious completion still returns. It is tempting because throttling genuinely mitigates abuse such as denial-of-service or runaway API consumption, where controlling request frequency is the goal rather than filtering harmful generated content.

• ✗

  Retraining the model on safe code only

  Why it's wrong here

  Retraining on safe code alone cannot guarantee refusal of every harmful prompt, since fine-tuning teaches patterns rather than enforcing runtime constraints; a novel phrasing may still elicit destructive output. It is tempting because retraining genuinely reduces unsafe completions during development, and would suit building a model whose baseline behaviour must be safer before deployment.