QuestionQ50

Securing AI systems

An attacker successfully carries out a denial-of-service (DoS) attack using the context window of an AI system. Thousands of characters are obfuscated and concealed behind an emoji. Which of the following techniques most effectively mitigates this type of attack?

Explanation

Prompt filtering is an input-side control that inspects requests before they reach the model. It can detect or normalize obfuscated input and block malicious or oversized payloads that would consume the context window and cause model denial of service. Microsoft’s Prompt Shields documentation identifies encoding attacks as adversarial input attacks and describes prompt shields as filters for generative-model inputs.

Learn more

Community Discussion

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