Cross-LLM Attacks COASP Practice Questions (Page 2)

Questions 6–10

1. 6application · medium

   A security team manages two separate LLM-based chat assistants: one for internal HR queries and one for customer support. Both assistants share a common system prompt template that includes a placeholder for 'current policy highlights.' An attacker discovers they can inject a malicious instruction into the HR assistant's output, which is then copied by employees into the customer support assistant as part of a troubleshooting query. The malicious instruction causes the customer support assistant to reveal its system prompt. What is the MOST direct reason this attack succeeded?

   • AThe HR assistant was not isolated from the customer support assistant at the network level.
   • BThe customer support assistant had a weaker architecture that could not resist prompt injection attacks.
   • CThe HR assistant's output was treated as untrusted input by the customer support assistant without proper sanitization.
   • DThe shared system prompt template allowed the injected instruction to be reused across both assistants.
   Reveal Solution Discussion Select an answer first
2. 7application · medium

   A security team wants to detect cross-LLM attacks in an environment where multiple LLM applications share data through a common message queue. The team has access to logs of all inputs and outputs for each LLM. What is the BEST approach to detect an indirect cross-LLM attack?

   • AAnalyze the message queue for patterns where an output from one LLM is followed by an anomalous behavior in another LLM.
   • BLook for any output from one LLM that contains the same text as the input to another LLM.
   • CReview the training data of all LLM applications for any malicious content.
   • DMonitor the network traffic between the LLM applications and the message queue.
   Reveal Solution Discussion Select an answer first
4. 8foundation · easy

   How can model outputs serve as a vector for cross-LLM attacks?

   • ABy consuming more computational resources than expected.
   • BBy being stored in a different file format than expected.
   • CBy containing malicious instructions that are executed when the output is used as input to another LLM.
   • DBy being displayed in a different language than the user's preference.
   Reveal Solution Discussion Select an answer first
5. 9application · medium

   A security operations center monitors two LLM systems that interact with each other: an email summarizer and a calendar assistant. The calendar assistant receives summaries from the email summarizer to create events. Analysts notice that the calendar assistant sometimes creates events with malicious links, but only when the email summarizer's output contains certain phrases. What detection method would BEST identify this as a cross-LLM attack?

   • ACorrelate the calendar assistant's anomalous behavior with the specific output patterns from the email summarizer.
   • BBlock all communication between the two LLM systems to prevent any potential attack.
   • CMonitor the email summarizer's output for any phrases that are not related to email content.
   • DReview the calendar assistant's training data for any malicious examples.
   Reveal Solution Discussion Select an answer first
6. 10foundation · easy

   Which scenario best illustrates a cross-LLM attack?

   • AAn attacker uses social engineering to trick a user into revealing their LLM API key.
   • BA malicious prompt is embedded in a document that is later processed by two different LLM systems, compromising both.
   • CAn attacker directly exploits a vulnerability in a single LLM's API endpoint.
   • DA denial-of-service attack overwhelms a single LLM's servers.
   Reveal Solution Discussion Select an answer first