LLM Jailbreaking COASP Practice Questions (Page 5)

Questions 21–25

1. 21expert · hard

   A security team is evaluating an LLM-based assistant that handles customer service. They have observed several jailbreak attempts. The team wants to implement a defense that is robust against multiple techniques, including role-playing, payload splitting, and encoding obfuscation. They also need to minimize false positives that would block legitimate user requests. Which approach best balances security and usability?

   • AOnly rely on the LLM's built-in safety alignment
   • BImplement layered defenses: input filtering, output filtering, and prompt engineering
   • CUse a single classifier to detect all jailbreak attempts
   • DDisable all external content processing
   Reveal Solution Discussion Select an answer first
2. 22foundation · easy

   Which of the following is a basic mitigation strategy for jailbreak attempts?

   • AIncreasing the model's temperature to make outputs more random
   • BRemoving all user input and only using system prompts
   • CDisabling the model's safety features to avoid false positives
   • DUsing input filtering to detect and block known jailbreak patterns
   Reveal Solution Discussion Select an answer first
4. 23foundation · easy

   How do token-level perturbations contribute to jailbreaking an LLM?

   • AThey encrypt the prompt so the model cannot understand it
   • BThey replace the model's system prompt with a malicious one
   • CThey introduce small changes to the input tokens that can cause the model to misinterpret the prompt and bypass safety
   • DThey alter the model's internal weights to disable safety mechanisms
   Reveal Solution Discussion Select an answer first
5. 24foundation · easy

   What is an adversarial suffix in the context of LLM jailbreaking?

   • AA technique that splits a harmful request into multiple parts
   • BA string of tokens appended to a prompt that is designed to trigger the model to ignore its safety guidelines
   • CA method of encoding the prompt in base64
   • DA way to change the model's persona to a fictional character
   Reveal Solution Discussion Select an answer first
6. 25foundation · easy

   How can indirect prompt injection be used to jailbreak an LLM?

   • ABy using a special token that resets the model's context
   • BBy modifying the model's training data to include harmful examples
   • CBy embedding malicious instructions in external content that the model processes, such as a webpage or document
   • DBy directly telling the model to ignore its guidelines in the user prompt
   Reveal Solution Discussion Select an answer first