1.3 Describe Vulnerabilities in AI/LLM Models Such as Prompt Injection, System Prompt Leakage, Vector and Embedding Weaknesses, and Supply Chain 350-701 Practice Questions (Page 3)

Questions 11–15

1. 11foundation · easy

   How can compromised training data introduce vulnerabilities into an LLM?

   • ABy increasing the model's accuracy on legitimate tasks.
   • BBy embedding biases or backdoors that cause the model to behave maliciously under certain conditions.
   • CBy making the model require more computational resources.
   • DBy causing the model to output random text.
   Reveal Solution Discussion Select an answer first
2. 12application · medium

   A security team is reviewing an AI-powered document search system that uses a vector database to retrieve relevant passages for an LLM to answer questions. An attacker submits a query that is slightly modified with a subtle, imperceptible text perturbation. The system retrieves a completely unrelated, sensitive document and includes it in the LLM's context. What is the most likely cause of this behavior?

   • AAn adversarial perturbation that exploits the embedding model's sensitivity to small input changes
   • BSystem prompt leakage exposing the retrieval logic
   • CA direct prompt injection in the user's query
   • DA supply chain compromise of the LLM model artifact
   Reveal Solution Discussion Select an answer first
4. 13foundation · easy

   What is an adversarial perturbation in the context of embedding models?

   • ASmall, carefully crafted changes to input data that cause the model to produce incorrect embeddings or outputs.
   • BA technique to compress embeddings to save storage space.
   • CA type of encryption used to protect embeddings from unauthorized access.
   • DA method to increase the accuracy of embeddings by adding more training data.
   Reveal Solution Discussion Select an answer first
5. 14expert · hard

   A security architect is reviewing the deployment of a fine-tuned LLM for medical diagnosis. The model was fine-tuned on a dataset that was assembled from multiple public sources. The architect is concerned about the integrity of the training data. Which of the following is the MOST comprehensive control to mitigate the risk of poisoned training data?

   • AImplement a data provenance and validation pipeline that checks the source and content of each training sample
   • BDeploy an output filter that blocks responses containing medical misinformation
   • CVerify the cryptographic hash of the final model artifact before deployment
   • DUse a different LLM provider that claims to have better security
   Reveal Solution Discussion Select an answer first
6. 15application · medium

   A company uses a vector database to store embeddings of its proprietary documents. An attacker gains write access to the database and inserts a malicious document embedding that is semantically close to legitimate queries about 'financial reports'. When users query the system, the malicious document is retrieved and its content is fed to the LLM, which then generates misleading answers. What type of attack is this?

   • ASupply chain compromise of the embedding model
   • BDirect prompt injection in the user's query
   • CSystem prompt leakage via the LLM
   • DData poisoning of the vector database
   Reveal Solution Discussion Select an answer first