Secret leakage
Also known as: secret leak · credential leak
Secret leakage is when a credential, key, or token slips into a prompt, log, or model output — and from there, out.
- A key in a log, copied into a prompt, echoed in output.
- A model summarizing a config that contains a secret.
- An agent writing a key to a shared file.
Prevention: keep secrets out of the model's context, filter logs, and treat any secret a model ever sees as compromised. Assume the worst.