Sensitive information poses risks to both the LLM itself and its application context. This includes personally identifiable information (PII), financial data, health records, confidential business information, security credentials, and legal documents. Additionally, proprietary models often have unique training methodologies and source code that are highly sensitive, particularly in closed or foundation models, making them targets for theft or misuse.

Remediation

• Limit the amount of sensitive information used during model training and inference to the bare minimum necessary for functionality.
• Mask or redact sensitive details before using them in prompts or as part of model inputs.
• Apply differential privacy to anonymise training data and prevent leakage of sensitive information through model outputs.