Keeping credentials out of prompts, logs and model context. Anything in the context window can end up in the output, in a log, or in a provider’s systems. Anything in a log can end up in a support ticket. The safe assumption is that a secret placed anywhere near a model will eventually be seen.
Technical view
Use a dedicated secret store with short-lived, narrowly scoped credentials issued per task, injected at the tool boundary rather than into model context, so the model requests an action rather than handling the key. Redact known secret patterns from logs and traces, rotate on suspicion, and scan trajectories for leaked material. Tokens for third-party connectors should be per-user and revocable rather than shared.
Related
Part of AI Terminology Index, section: Security and Privacy.