LLM 프롬프트의 개인정보 보호와 데이터 보호

Whatever you put in a prompt is sent to a third-party model provider. If it contains names, emails, health, or financial data, you may be exposing — and even logging — personal information.

Personally Identifiable Information includes anything that identifies a person: names, emails, phone numbers, addresses, IDs, and combinations that together identify someone.

Laws govern how you handle personal data:

• GDPR (EU): consent, purpose limitation, right to deletion
• CCPA (California): disclosure and opt-out
• HIPAA (US health): protected health information

The golden rule: send the model only what it truly needs. Strip fields irrelevant to the task before they ever reach the prompt.

Detect and mask PII patterns in user input before it goes to the model, then optionally restore it in the final output.

Replace each PII value with a placeholder token, keep the mapping locally, and swap real values back only after the model responds — the provider never sees raw data.

Read the provider terms: does your data train their models? Is it logged, and for how long? Enterprise tiers often offer zero-retention and no-training guarantees.

Under GDPR users can demand deletion. Design your system so prompts and outputs are not silently retained where you cannot erase them on request.

For the most sensitive data, run an open model on your own infrastructure so prompts never leave your network — trading some capability for full data control.

Tell users when AI processes their data and why. Clear consent is both a legal requirement and a trust builder.

You need logs for debugging, but logs of raw prompts can themselves become a breach. Log redacted prompts and restrict access to anyone who can view raw data.

You learned to protect privacy in prompts: identify PII, follow GDPR/CCPA/HIPAA, apply data minimization and redaction, use reversible masking, check provider retention policies, support deletion rights, consider local models, and log safely.