Keep Secrets Away From Agents
A secret, such as a credential, a token or an API key, is set up so only the running app can read it, as far from the agent as the setup allows.
Why
What an agent reads can end up in a chat, a log or a commit, and a secret that leaks there is out of the director's hands.
How To Apply
- A secret lives in the environment the app runs in, such as the host's or the CI's secret store, never in a file or a message the agent reads.
- A
.env*file with secrets is avoided, even where the director points to one. Configuration that is no secret can live in the project. - When a task needs a secret, the agent sets up where the app reads it, and the director puts the value there.
- A secret that reaches the chat anyway is never written to a log, a record or a commit, the request log included.
- The keychain is never touched (Never Access The Keychain).