Cross-LLM Attacks COASP Practice Questions (Page 6)

Questions 26–30

1. 26application · medium

   A security team wants to detect cross-LLM attacks in an environment where multiple LLM systems exchange outputs. Which monitoring approach is most likely to identify an ongoing cross-LLM attack?

   • AMonitor the CPU and memory usage of the servers hosting the LLM systems.
   • BReview the training data of each LLM for poisoned samples.
   • CTrack the flow of outputs between LLM systems and flag when a model's output contains instructions that are subsequently executed by another model.
   • DCheck the network firewall logs for unusual outbound connections from the LLM servers.
   Reveal Solution Discussion Select an answer first
2. 27foundation · easy

   Which method is used to detect cross-LLM attacks?

   • AChecking the physical temperature of the servers.
   • BCounting the number of users logged into the system.
   • CMonitoring for anomalous patterns across model interactions.
   • DReviewing the marketing materials of the LLM providers.
   Reveal Solution Discussion Select an answer first
4. 28application · medium

   A startup uses a large public LLM to generate synthetic customer support dialogues, which are then used to fine-tune a smaller internal LLM. An attacker poisons the public LLM's outputs so that the fine-tuned internal model becomes biased toward revealing customer payment details. What is the primary attack vector?

   • AShared prompt attack, because the same prompt was used for both models.
   • BOutput filtering failure, because the internal model failed to filter its own outputs.
   • CIndirect prompt injection, because the attacker injected a prompt into the public LLM.
   • DModel-to-model data poisoning, because the poisoned outputs from the public LLM contaminated the fine-tuning data of the internal model.
   Reveal Solution Discussion Select an answer first
5. 29application · medium

   A company uses an LLM to generate meeting summaries and another LLM to create action items from those summaries. An attacker embeds a hidden instruction in a meeting summary that, when processed by the action-item LLM, causes it to output a malicious URL. What is this attack called?

   • AAn indirect cross-LLM attack, where malicious content in one LLM's output compromises another LLM.
   • BA model inversion attack that extracts training data from the action-item LLM.
   • CA data poisoning attack on the summarization LLM's training data.
   • DA direct prompt injection attack on the action-item LLM.
   Reveal Solution Discussion Select an answer first
6. 30expert · hard

   A company 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 is considering two mitigations: (1) filter the LLM's output before using it as training data, and (2) use a different LLM provider for the smaller model's training data. Which mitigation is MORE effective in preventing model-to-model data poisoning?

   • ANeither is effective because the attack is already in the training data and cannot be removed.
   • BBoth are equally effective because they both prevent the malicious content from entering the training data.
   • CUsing a different LLM provider is more effective because it ensures the training data is generated by a different model.
   • DFiltering the LLM's output is more effective because it directly removes malicious content before it contaminates the training data.
   Reveal Solution Discussion Select an answer first