Before an LLM request leaves your app, inspect what it contains

Описание и применение

Sether позволяет заменять чувствительные данные в запросах к LLM токенами, уменьшая риск утечки информации, но не гарантирует полное соответствие требованиям конфиденциальности.

Задача

Sensitive data can leak from application to LLMs when requests are sent without proper redaction, risking privacy violations.

Результат

Sether enables controlled redaction of sensitive data in LLM prompts, allowing developers to verify what data leaves the application and to restore tokens in responses, though it does not guarantee full compliance.

Рабочий процесс

• 1. Create fictional messages with task instructions and sensitive values.
• 2. Run Sether to redact sensitive values, producing a tokenized prompt.
• 3. Inspect the redacted request, counting missed values and unnecessary replacements.
• 4. Send the redacted prompt to the LLM and evaluate task quality.
• 5. Receive the response, check for preserved tokens, and restore original values using the mapping.
• 6. Test edge cases: altered tokens, missing mappings, and streaming where values split across chunks.

Доказательства

I am building Sether around that boundary. It is an open-source library that replaces detected sensitive values with stable tokens before text goes to an LLM.

Источник

The useful question is not “did we add a privacy tool?” It is “what actually left the application?”

Источник

A practical evaluation has four parts: ...

Источник

Streaming adds another case: a value may be split across chunks. Include that in your tests rather than relying only on complete strings.

Источник

Sether is not a promise that every sensitive value will be detected, and installing it does not establish regulatory compliance.

Источник