Ai Safety Secrets And Governance

A Droid reads an env var holding a token to call an API and prints it into the run transcript while debugging.

Reference secret values without printing them: never echo a token/key into logs, the transcript, commit messages, or error output. Redact secret-shaped values before logging. Assume the transcript is visible to people who should not see the secret.

An auditor asks for evidence of Factory's SOC 2 posture as a sub-processor. The operator quotes a marketing line instead of the report.

Pull the actual current attestation from Factory's trust center (under NDA if required), confirm the audit period covers the relevant window, and map controls to the operator's own program — do not cite a marketing badge. Treat the report snapshot as time-bound. [REQUIRES-VERIFICATION] for current …

A coding task could be 'finished faster' by editing the running production system directly instead of going through the deploy pipeline. The Droid does it.

Make changes through the reviewed code path (branch → PR → CI → deploy), never by editing production directly. The prod-write boundary is a hard line; a shortcut that mutates live systems outside the pipeline is prohibited even when faster.