When robust input controls are not practical on a large language model (LLM) to prevent prompt injection attacks from external threats, which of the following would be the BEST compensating control…

Question

Options

• AReview and annotate the AI system's outputs
• BImplement identity and access management (IAM)
• CConduct human reviews of the AI system's inputs
• DFine-tune the system to validate the AI system's inputs

• A
  76% (28)
• B
  8% (3)
• C
  14% (5)
• D
  3% (1)

Explanation

When prompt injection cannot be reliably prevented at the input layer, a compensating control shifts focus to detecting and containing successful attacks at the output layer. Reviewing and annotating the LLM's outputs allows human or automated review to identify anomalous, harmful, policy-violating, or adversarially influenced responses before they are delivered or acted upon - limiting the blast radius of a successful injection. IAM (B) manages access but doesn't address injected content within authorized sessions. Human review of inputs (C) is costly, slow, and impractical at scale for LLMs. Fine-tuning to validate inputs (D) is essentially an input control, not a compensating control when input controls are explicitly deemed impractical.

No community discussion yet for this question.