Report a vulnerability to security@latere.ai. Do not open a public issue for one. You will hear back within three business days, and a fix for a high severity issue ships within thirty days. Credit in the release notes on request.

Fixes go to the two most recent minor release series. There is no release yet; the first one is v0.1.0.

What Cella protects, against whom, and how each threat is answered is written down in the threat model, where every control names the test that proves it, so a reviewer can check the design rather than take it on faith. The commitments the design makes: every /v1 request carries a token from an issuer the operator listed or one the control plane minted, and nothing acts before the authorizer has decided, with a refused object answering exactly as a missing one; a decision the authorizer cannot give is a refusal, never an allow; a sandbox reaches the network only through its egress gateway, which admits what its manifest's mode and host lists allow, on every environment that enforces egress; a workload token identifies one sandbox and is refused by the control plane once that sandbox is gone.

Dependencies are checked for known vulnerabilities on every push. A release will carry an SPDX bill of materials for the module graph and one per image, cosign signatures over the images and the checksums, and an SBOM and a build provenance attestation per image, so gh attestation verify answers for the image you are about to run.