Cross-LLM Attacks COASP Practice Questions (Page 8)

Questions 36–37

1. 36expert · hard

   A company uses a public LLM to generate synthetic data for fine-tuning an internal LLM. An attacker poisons the public LLM's outputs to introduce a backdoor in the internal model. The security team wants to mitigate this risk. Which mitigation is most effective?

   • AValidate and sanitize the synthetic data before using it for fine-tuning.
   • BDisable the internal model's ability to output sensitive data.
   • CUse a different public LLM that is known to be secure.
   • DIncrease the amount of synthetic data to dilute the poisoned samples.
   Reveal Solution Discussion Select an answer first
2. 37application · medium

   A company uses LLM-A to generate meeting summaries that are automatically emailed to employees. LLM-B is an internal knowledge assistant that indexes those emails and answers questions based on them. An attacker crafts a meeting summary that, when read by LLM-B, causes it to exfiltrate sensitive data from the knowledge base. What type of attack is this?

   • AIndirect cross-LLM attack, because the malicious content was embedded in LLM-A's output and later processed by LLM-B.
   • BModel-to-model data poisoning, because LLM-A's output contaminated LLM-B's training data.
   • CA denial-of-service attack, because the malicious summary degraded LLM-B's performance.
   • DDirect prompt injection, because the attacker directly sent a prompt to LLM-B.
   Reveal Solution Discussion Select an answer first