Cross-LLM Attacks COASP Practice Questions (Page 5)

Questions 21–25

1. 21foundation · easy

   What is the purpose of output filtering as a mitigation for cross-LLM attacks?

   • ATo change the language of the LLM's responses.
   • BTo increase the speed of LLM responses.
   • CTo remove malicious content from LLM outputs before they are used elsewhere.
   • DTo make the LLM's responses more verbose.
   Reveal Solution Discussion Select an answer first
2. 22foundation · easy

   What is the defining characteristic of a cross-LLM attack?

   • AAn attack that exploits a vulnerability in the LLM's underlying hardware.
   • BAn attack that only affects the specific LLM instance it was originally targeted at.
   • CAn attack that propagates or transfers between different LLM systems.
   • DAn attack that requires physical access to the LLM's hosting infrastructure.
   Reveal Solution Discussion Select an answer first
4. 23expert · hard

   A company has a pipeline where LLM-A generates reports that are automatically sent to LLM-B for summarization. An attacker poisons LLM-A's outputs to include a hidden instruction that causes LLM-B to leak sensitive data. The security team wants to mitigate this without disrupting the pipeline. Which mitigation is most effective?

   • AIncrease the temperature of LLM-B to make its outputs less predictable.
   • BUse a different LLM for LLM-B that is not susceptible to prompt injection.
   • CDisable LLM-B's access to the sensitive data.
   • DSanitize LLM-A's outputs by removing any text that looks like an instruction before sending them to LLM-B.
   Reveal Solution Discussion Select an answer first
5. 24expert · hard

   An offensive AI specialist is testing a prompt injection payload against multiple LLMs from different vendors. The payload works on the first LLM but fails on the second. The specialist wants to modify the payload to increase the chance of success across all models. What is the MOST effective modification?

   • AMake the payload more generic by removing vendor-specific references and using common instruction patterns.
   • BUse a payload that exploits a known vulnerability in the second model's architecture.
   • CEncode the payload in a format specific to the first model's tokenizer to ensure it is not filtered.
   • DMake the payload longer and more detailed to overwhelm the second model's safety mechanisms.
   Reveal Solution Discussion Select an answer first
6. 25foundation · easy

   Which of the following is a mitigation strategy to prevent cross-LLM attack spread?

   • ASharing prompts across all LLM systems for consistency.
   • BIsolating LLM systems from each other.
   • CIncreasing the number of LLM systems in use.
   • DDisabling all logging of model interactions.
   Reveal Solution Discussion Select an answer first