Which of the following BEST reduces the risk of exposing sensitive data through the output of large language models (LLMs) in applications?

The most effective method to prevent large language models (LLMs) from exposing sensitive data in their output is to ensure they are never exposed to it in the first place. Data sanitization techniques, such as masking, redaction, or tokenization, are applied to the training data and user prompts to remove or anonymize personally identifiable information (PII) and other sensitive data before the LLM processes it. By cleaning the input, the model does not learn the sensitive information and therefore cannot inadvertently reproduce or "regurgitate" it in its responses. This is a direct, preventative control that addresses the root cause of the risk.

A. Encrypting data in transit and at rest protects data during storage and transmission but not during processing. The LLM requires decrypted data to function, at which point it can be exposed.

B. Conducting adversarial testing is a validation technique used to identify if an LLM is vulnerable to leaking sensitive data, rather than a preventative control that reduces the risk itself.

D. Enforcing least privilege access restricts who can interact with the LLM or its underlying data, but it does not prevent an authorized user or process from using sensitive data that the model can then leak.

1. ISACA

Artificial Intelligence Audit Toolkit

2023. In the "AI Governance and Risk Management" domain

Control Objective A&G-05: "Data Privacy and Protection" emphasizes implementing controls to protect sensitive data throughout the AI lifecycle. This includes the data preprocessing stage

where sanitization is a key activity to prevent PII from being ingested by the model.

2. ISACA

Auditing Artificial Intelligence White Paper

2023. Page 12

under the "Data Acquisition and Preparation" section

states that a key audit step is to "verify that controls are in place to ensure that PII and other sensitive data are removed or anonymized from data sets prior to their use in AI/ML models." This directly supports sanitization as the primary control.

3. Mee

T.

et al. A Survey on Privacy in Large Language Models: Attacks

Defenses

and Future Directions. arXiv

2023. Section 4.1

"Data Pre-processing

" identifies data sanitization and anonymization as primary defense mechanisms against privacy attacks. It states

"The most straightforward defense is to remove sensitive information from the pre-training data." (DOI: https://doi.org/10.48550/arXiv.2312.06717).