CVE-2026-63343

This CVE exposes a pattern you need to recognize: the same symlink-escape vulnerability recurring across different metadata file paths in Incus because the original patches were path-specific rather than architectural. The earlier fixes for `exec-output` and `templates/` used two different mechanisms — `Lstat` rejection for one path and `os.OpenRoot` confinement for another — proving the team understood the vulnerability class but applied that understanding selectively rather than broadly. `metadata.yaml` simply wasn't in scope when those patches shipped. The critical question for version 7.3.0: does the fix apply `os.OpenRoot` confinement to the entire image metadata parsing pipeline, or does it add only a `metadata.yaml` exception to existing path handling? If it's the latter, expect follow-up CVEs — the same vulnerability class will surface in other metadata paths that haven't been audited. Check the commit message: language like "apply os.OpenRoot to all metadata file parsing" signals architectural remediation; language like "add metadata.yaml to existing scope" signals another path-specific patch. What to do: audit every path parameter in your image specification parsing code that accepts user-controlled file references. The pattern isn't sophisticated — it's the same symlink-escape mechanism applied wherever a metadata file path wasn't hardened during the earlier patch cycle. Prioritize paths that haven't been touched in recent commits; forgotten code paths are where these vulnerabilities survive between security sprints. The deeper failure here is systemic. CVE descriptions encode path-level knowledge, not class-level knowledge. This means other projects parsing OCI-style image metadata — Buildah, Podman, custom import pipelines — likely have the same exposure and may not realize it because the earlier Incus CVEs didn't communicate the vulnerability class. If you're maintaining any container image parsing code, treat the CVE pattern as a class-level alert, not an Incus-specific incident.

Reviewed through automated stages and approved by a human before publication.