Cross-LLM Attacks COASP Practice Questions (Page 4)

Questions 16–20

1. 16expert · hard

   A company uses an LLM to generate meeting summaries and another LLM to create action items. An attacker embeds a hidden instruction in a meeting summary that causes the action-item LLM to output a malicious link. The security team wants to prevent this attack without blocking the legitimate flow of summaries to the action-item LLM. Which mitigation is MOST appropriate?

   • ABlock all communication between the two LLMs to prevent the attack.
   • BRequire human review of every meeting summary before it is used by the action-item LLM.
   • CApply output filtering on the meeting summary LLM to remove any instruction-like content before it is passed to the action-item LLM.
   • DUse a different LLM provider for the action-item LLM to ensure it is not affected by the same attack.
   Reveal Solution Discussion Select an answer first
2. 17foundation · easy

   How can poisoned outputs from one LLM contaminate another LLM?

   • ABy being included in the training dataset used to fine-tune the second LLM.
   • BBy increasing the second LLM's response time.
   • CBy changing the color scheme of the second LLM's interface.
   • DBy causing a temporary network slowdown in the second LLM's servers.
   Reveal Solution Discussion Select an answer first
4. 18expert · hard

   A data science team uses an LLM to generate synthetic training data for a smaller model. The LLM is compromised by a prompt injection attack that causes it to generate biased examples. The team discovers the attack after the smaller model has already been trained. What is the BEST course of action?

   • ARetrain the smaller model using only filtered and validated data from the LLM.
   • BContinue using the smaller model but monitor its outputs for biased behavior.
   • CApply output filtering to the smaller model's outputs to remove any biased content.
   • DReplace the smaller model with a model trained on data from a different LLM provider.
   Reveal Solution Discussion Select an answer first
5. 19application · medium

   An organization has two LLM systems: one that generates reports and another that processes those reports to extract action items. They suspect that malicious content in the reports could compromise the action-item extractor. Which mitigation is most effective at preventing this cross-LLM attack?

   • AIncrease the temperature setting of the action-item extractor to make its outputs more random.
   • BDisable the action-item extractor's ability to access the internet.
   • CUse the same LLM for both tasks to reduce the number of systems.
   • DSanitize the report text before feeding it to the action-item extractor by stripping or neutralizing known prompt-injection patterns.
   Reveal Solution Discussion Select an answer first
6. 20expert · hard

   A company has a pipeline where LLM-A generates summaries that are automatically sent to LLM-B for further processing. They want to prevent cross-LLM attacks but also need to maintain the pipeline's efficiency. Which mitigation is most effective?

   • ASanitize LLM-A's outputs by removing instruction-like patterns before sending them to LLM-B.
   • BIsolate LLM-A and LLM-B by using a human-in-the-loop review of LLM-A's outputs before they are sent to LLM-B.
   • CDisable LLM-B's ability to access sensitive data.
   • DUse a different LLM for LLM-B that is not susceptible to prompt injection.
   Reveal Solution Discussion Select an answer first