probaci executes third-party container images, runs CI tools over repository content, and handles tokens. Hardening spans four layers. To report a vulnerability, see SECURITY.md.
The original bash tool used eval; probaci eliminates that class entirely.
Tools run via exec with argument vectors only — config values are never
concatenated into a shell string. Config is schema-validated (unknown fields
rejected), and target paths are resolved and confined to the repository root.
Every brokered run defaults to:
--rm --cap-drop ALL --security-opt no-new-privileges
--read-only --tmpfs /tmp (writable scratch only)
--user <host uid:gid> (non-root, on Linux)
--pids-limit / --memory / --cpus
--network none (offline unless the tool needs the network)
-v <repo>:/workspace:ro (read-only repo; clean-clone uses an isolated rw export)
Backends: Docker (default), rootless Docker, and Podman — choose in
config or --backend.
Nested-Docker runners (act, gitlab-ci-local) may need the container socket.
probaci never mounts it for its own tool runs; enabling it for those runners
is opt-in (docker.allow_socket_mount) and is best done on the rootless/Podman
backend.
- Passed via env file or stdin, never argv.
- Secret files live
0600under~/.config/probaci/secrets/and are git-ignored by default. - A central redactor scrubs known secret values and token-shaped strings
from stdout,
--json, and the rotating log file.
- Images are referenced by digest where pinned (tamper-evident).
- cosign verification is advisory by default (verify + warn);
strictmode refuses unverified images unless allow-listed. - A private
registry_mirrorcan be configured.
Releases are reproducible and carry SHA-256 checksums, cosign signatures,
an SBOM, and SLSA provenance, published via OIDC trusted publishing. CI
runs govulncheck, golangci-lint, and gitleaks, with all Actions pinned by
commit SHA and permissions: contents: read by default.