Please do not open a public issue for a security problem.

Use GitHub's private reporting — Security → Report a vulnerability — or email the maintainer listed on the repository profile.

You will get an acknowledgement within a few days. This is a small project with one maintainer, so please be patient; if a fix takes longer than that, you will be told where it stands rather than left waiting.

upkeep holds cloud credentials in memory and writes changes to production infrastructure, so the interesting failures are about disclosure and about doing something the operator did not review.

Especially in scope:

• A value reaching an output. A plan, -json, a saved plan, an error, or an import that prints or writes a secret. Every one of these is meant to name where a value came from and never the value itself.
• A reference escaping into a subprocess. valueFrom references are untrusted input; commands run without a shell and references are validated before anything runs. A way around either is a vulnerability.
• Applying something that was not planned. A saved plan that applies after the world moved, or an action executing that the plan did not list.
• A MANUAL action executing. Those must never have a Do.
• Credential handling. Anything that writes a borrowed CLI session to disk, logs a token, or leaks one through an error or a %v.

• A hostile config file. A config can name any service or account you have access to, and valueEnv can name any variable in your environment. Review a config you did not write, as you would any script. This is documented, not a bug.
• A mis-named value. The classifier that refuses a literal under a credential-shaped name reads names, not contents. It will not catch a secret in a variable called SETTINGS_BLOB. Documented limitation.
• The access a borrowed session already has. upkeep trusts what render login, wrangler login and neonctl auth stored, and has exactly their permissions.
• Anything requiring an attacker to already have your shell.

The latest release. This project is young enough that backporting to older tags is not sensible yet; if that changes, this section will say so.

A security fix gets a release of its own, notes saying what was exposed and under what conditions, and a test that fails without the fix. If the problem was a disclosure, the notes say plainly what could have been disclosed rather than describing it as hardening.