sk

Agent Security FAQ

1) Why avoid putting secrets directly in prompts?

Prompts can be logged, cached, exported, or reused by downstream tools. Keep secrets outside prompt text and inject them only at execution time.

2) What is the safest way to provide credentials to AI workflows?

Use runtime injection with environment variables and stdin-based flows. With sk, store once and fetch only in the process that needs the secret.

3) How should teams split credentials across environments?

Use separate keys per environment (dev/staging/prod) and per automation surface (local agent, CI, production worker).

4) What should be redacted before sharing logs or traces?

Redact API keys, bearer tokens, session ids, internal URLs, user identifiers, and tool outputs that could reveal secret material.

5) What is a practical rotation policy?

Rotate on schedule (for example monthly), rotate after role changes, and rotate immediately after suspected exposure. Revoke old keys in the same runbook.

Quick companion

• Security checklist card: ./agent-security-checklist.md

This site is open source. Improve this page.