Report privately through GitHub's security advisories: https://github.com/yannickYamo/atelier/security/advisories/new

Please do not open a public issue for a vulnerability. There is no bug bounty. I aim to acknowledge within seven days and to disclose within ninety days of a fix, or sooner by agreement.

The latest release only. This project is pre-1.0 and there is no backport branch.

• It reads a corpus you point it at. Discovery sends that material to whichever inference provider you configured.
• Writing and checking a draft send text to the same provider. invoke sends the task, the compiled skill and any material you bind to the writer model. With a key set, verify and invoke also send the draft and that material to a small claim-reading model (claude-haiku-4-5 by default, or ATELIER_CLAIMS_MODEL), and the taste reader sends the draft and the rules it reads. --claims pattern, --no-taste and atelier verify without a key keep everything on the machine. Nothing goes anywhere but the provider you configured.
• It stores state locally, under $ATELIER_DATA or ~/.atelier: standards, compiled packages, the ratification ledger, invocation records. No telemetry, no account, no network calls except to your configured provider.
• It reads one credential from the environment, named by --api-key-env and defaulting to your provider's convention. It is never written to disk and never included in a record.
• It writes into directories you name, including agent-host skill directories. revert restores what a build overwrote; rollback moves between versions Atelier itself minted.

If you find a path where a credential, a corpus, or held-out material leaks somewhere the above does not describe, that is a vulnerability and I want to know.