Your AI coding agent can access the terminal, environment variables, config files, and even that .env file you meant to delete. It uses your credentials to run npm install and includes the command output in its context. Attackers are aware of this. In April 2026 (Mini Shai-Hulud), a compromised npm package delivered a payload aimed at authenticated AI coding assistants.

Most practices for handling developer secrets were created when only humans worked on code. We'll move fast through five areas where the old habits break down in the agent era:

- Local development: why that .env file is a bigger liability than you think, and what replaces it.
- CI/CD: how to stop storing cloud keys in GitHub Actions entirely.
- Tokens: the case for short-lived, narrowly scoped credentials, and why your personal access token shouldn't be doing the job.
- AI agents: what to do when your agent can read everything on your machine, and how to give it far less to leak.
- Safety nets: the guardrails that catch a secret before it ships, and the plan you'll want ready for when one slips through anyway.

You’ll get a checklist you can use in your own repo right away, plus one simple rule for working with agents: if your agent can read a secret, you should assume it can leak it.