Multiple choice :

Q$1.$ What mitigation measure is most effective against indirect prompt injections in LLMs$?$

$1-$ Direct prompt injections are more difficult to implement than indirect prompt injections.

$2-$ Increasing computational resources for the LLM$.$

$3-$ Implementing API token limits for backend system interactions.

$4-$ Segregating trusted and untrusted input sources.

$5-$ Expanding the LLM$\text{'}$s training data set $.$

Q$2.$ Which outcome is a direct result of Insecure Output Handling in LLMs$?$

$1-$ Reduced inference time on computational tasks.

$2-$ Potential execution of unauthorized code on backend systems.

$3-$ Increased accuracy of language model predictions.

$4-$ Enhanced data privacy and security.

Q$3.$ Considering the mitigation strategies for prompt injection vulnerabilities, which approach is least likely to be effective?

$1-$ Applying the principle of least privilege to LLM operations.

$2-$ Increasing the number of data sources the LLM can access.

$3-$ Manual monitoring of LLM inputs and outputs.

$4-$ Implementing human oversight for high$-$risk actions.

Q$4.$ In the context of LLMs$,$ what is the primary threat posed by Training Data Poisoning?

$1-$ It decreases the computational speed of LLMs$.$

$2-$ It causes the LLM to produce biased or incorrect outputs.

$3-$ It enhances the transparency of the LLM$\text{'}$s training process.

$4-$ It improves the LLM$\text{'}$s ability to handle large datasets.

Q$5.$ Which of these is NOT a recommended practice to mitigate Model Denial of Service attacks in LLMS$?$

$1-$ Employing machine learning models to automatically filter out potential DoS inputs.

$2-$ Enforcing strict APl rate limits$.$

$3-$ Reducing the frequency of model retraining.

$4-$ Monitoring for abnormal resource consumption patterns.