QuestionQ25

VerifiedManaging Risks of Generative AI in Software Testing

Match each Generative AI model attack-vector type (1–4) to its corresponding software-testing example (A–D).

Attack Vector

  1. Data exfiltration
  2. Request manipulation
  3. Data poisoning
  4. Malicious code generation

A) An attacker submits a carefully crafted input to an LLM, causing it to unintentionally disclose excerpts of confidential training data, such as internal API keys.

B) When asked to assess the quality of an AI-generated test report, a malicious tester repeatedly provides inaccurate or deceptive evaluations, degrading the model’s future performance.

C) A user deliberately places a subtly modified image in a multimodal prompt, leading the LLM to misunderstand the context and produce irrelevant or incorrect acceptance criteria for a test task.

D) An adversary crafts a prompt to force the LLM to generate executable code that, when executed, tries to create an unauthorized communication channel to an externally controlled server.

Explanation

Data exfiltration involves revealing confidential information, including sensitive training data. Request manipulation includes crafted multimodal inputs that alter model behavior or produce misleading output. Data poisoning manipulates learning or evaluation data, impairing future model performance. Malicious code generation entails inducing a model to generate executable code intended for unauthorized or harmful actions.

Learn more

Community Discussion

No comments yet. Be the first to start the discussion!