feat(build): publish OCI SBOM and provenance attestations (#2836)

* feat(build): embed auditable Rust dependency metadata

Signed-off-by: Adrien Langou <alangou@nvidia.com>

* feat(build): publish OCI SBOM and provenance attestations

Signed-off-by: Adrien Langou <alangou@nvidia.com>

---------

Signed-off-by: Adrien Langou <alangou@nvidia.com>

.agents/skills/sbom/SKILL.md

@@ -13,13 +13,28 @@ The OpenShell SBOM tooling produces source-tree CycloneDX JSON SBOMs using Syft,
 
 SBOMs are **release artifacts only** -- they are generated on demand and not committed to the repository. Output lands in `deploy/sbom/output/` (gitignored).
 
-Release Dev and Release Tag image builds separately embed cargo-auditable metadata in the staged gateway and supervisor binaries. This metadata describes the binary's Rust dependency graph and lets Syft discover Cargo packages from the binary itself. It is not a complete image SBOM and is not an OCI SBOM attestation; publishing such an attestation remains separate work.
+Pushed gateway and supervisor images carry an SPDX SBOM and minimal SLSA provenance as OCI attestations. Branch E2E, Release Dev, and Release Tag image binaries embed cargo-auditable metadata, so their image SBOMs include linked Rust crates.
 
 ## Prerequisites
 
 - `mise install` has been run (installs Syft and other tools)
 - The repository is checked out at the root
 
+## Inspecting an Image SBOM
+
+BuildKit uses its default Syft scanner and attaches one SPDX document per platform. Read one without pulling the image:
+
+```bash
+docker buildx imagetools inspect ghcr.io/nvidia/openshell/gateway:latest \
+  --format '{{ json (index .SBOM "linux/amd64").SPDX }}'
+```
+
+Validate the final attestation, requiring a Cargo package for an auditable image:
+
+```bash
+tasks/scripts/verify-image-sbom.sh ghcr.io/nvidia/openshell/gateway:latest --require-cargo
+```
+
 ## Inspecting an Auditable Image Binary
 
 Opt into auditable metadata when staging a local image binary:

.github/workflows/branch-e2e.yml

@@ -86,6 +86,7 @@ jobs:
     with:
       component: gateway
       image-tag: ${{ github.sha }}
+      auditable: true
 
   build-supervisor:
     needs: [pr_metadata]
@@ -97,6 +98,7 @@ jobs:
     with:
       component: supervisor
       image-tag: ${{ github.sha }}
+      auditable: true
 
   build-cli:
     needs: [pr_metadata]

.github/workflows/docker-build.yml

@@ -294,6 +294,10 @@ jobs:
       volumes:
         - /var/run/docker.sock:/var/run/docker.sock
     steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+        with:
+          ref: ${{ inputs['checkout-ref'] || github.sha }}
+
       - name: Log in to GHCR
         run: echo "${{ secrets.GITHUB_TOKEN }}" | docker login ghcr.io -u "${{ github.actor }}" --password-stdin
 
@@ -309,3 +313,14 @@ jobs:
             --prefer-index=false \
             -t "${image}:${{ needs.resolve.outputs.image_tag_base }}" \
             "${refs[@]}"
+
+      - name: Verify merged manifest SBOM attestation
+        env:
+          IMAGE_REF: ghcr.io/nvidia/openshell/${{ inputs.component }}:${{ needs.resolve.outputs.image_tag_base }}
+          REQUIRE_CARGO: ${{ inputs.auditable }}
+        run: |
+          args=("${IMAGE_REF}")
+          if [[ "${REQUIRE_CARGO}" == "true" ]]; then
+            args+=(--require-cargo)
+          fi
+          tasks/scripts/verify-image-sbom.sh "${args[@]}"

.github/workflows/rust-native-build.yml

@@ -40,6 +40,8 @@ on:
         required: false
         type: string
         default: ""
+      # Branch E2E and release gateway/supervisor image builds opt in.
+      # CLI images and standalone glibc-static supervisor validation keep false.
       auditable:
         description: "Embed cargo-auditable dependency metadata in the binary"
         required: false

architecture/build.md

@@ -131,27 +131,39 @@ step via the `rust-native-build.yml` workflow (per-architecture, per-component)
 and uploads the result as an artifact that the image build job downloads back
 into the staging directory before running Buildx.
 
-Gateway and supervisor binaries staged into Release Dev and Release Tag images
-are compiled through `cargo auditable` (pinned in `mise.toml`), which embeds a
-`.dep-v0` section describing the Rust dependencies actually compiled into the
-binary. That section holds data rather than symbols, so it survives the
+Gateway and supervisor binaries staged into branch E2E, Release Dev, and Release
+Tag images are compiled through `cargo auditable` (pinned in `mise.toml`), which
+embeds a `.dep-v0` section describing the Rust dependencies actually compiled
+into the binary. That section holds data rather than symbols, so it survives the
 workspace's `strip = true` release profile, and Syft can catalog the crates
 present in image binaries instead of inferring them from the source tree. This
 is a different artifact from the source SBOM produced by `syft dir:.` in
-`tasks/sbom.toml`, which describes the checkout, and from an OCI SBOM
-attestation, which remains out of scope.
+`tasks/sbom.toml`, which describes the checkout, and from the image SBOM
+attestation below, which describes a published image.
 
-Only release image builds are auditable. `docker-build.yml` and
-`rust-native-build.yml` take an `auditable` input that defaults to false, so PR
-and E2E image builds and standalone release artifacts remain non-auditable. The
-CI image gains the pinned `cargo-auditable` tool through `mise install --locked`
-but ships no auditable OpenShell binary of its own. Local staging opts in with
+`docker-build.yml` and `rust-native-build.yml` take an `auditable` input that
+defaults to false. The branch E2E gateway and supervisor image builds set it to
+true so E2E exercises release-parity image binaries; other PR image builds and
+standalone release artifacts stay non-auditable. The CI image gains the pinned
+`cargo-auditable` tool through `mise install --locked` but ships no auditable
+OpenShell binary of its own. Local staging opts in with
 `OPENSHELL_AUDITABLE=1`. sccache's `RUSTC_WRAPPER` is unset only around auditable
 builds, because it would otherwise wrap `cargo-auditable`'s workspace wrapper and
 be misidentified as `rustc`. Auditable builds are verified by scanning the built
 binary with Syft and requiring at least one decoded Cargo package; the check runs
 only for those builds.
 
+Pushed Docker images carry minimal SLSA provenance and a per-platform SPDX SBOM
+generated by BuildKit's default Syft scanner. The registry exporter uses OCI
+media types and `oci-artifact=true`, so each attestation identifies its subject.
+GHCR exposes these through the image index because it has no referrers API.
+
+Attestations require a registry-backed image index. Local builds therefore keep
+`--provenance=false`, and Podman builds carry neither attestation.
+`tasks/scripts/verify-image-sbom.sh` verifies the merged multi-arch tag and runs
+with `--require-cargo` for auditable builds, so those attestations must also
+contain Cargo packages.
+
 Runtime layout:
 
 - **Gateway**: `gcr.io/distroless/cc-debian13:nonroot` base, GNU-linked binary at

docs/security/verifying-images.mdx

@@ -0,0 +1,36 @@
+---
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+title: "Verify the Contents of Published OpenShell Images"
+sidebar-title: "Verify Image Contents"
+slug: "security/verify-image-contents"
+description: "Read the SBOM attestation attached to published gateway and supervisor images to audit what each image contains."
+keywords: "Generative AI, Cybersecurity, Supply Chain, SBOM, Container Images"
+position: 2
+---
+
+Published gateway and supervisor images carry one SPDX SBOM per platform as OCI attestations.
+
+## Inspect an Image
+
+Read a platform's document without pulling the image:
+
+```shell
+docker buildx imagetools inspect ghcr.io/nvidia/openshell/gateway:latest --format '{{ json (index .SBOM "linux/amd64").SPDX }}'
+```
+
+List the packages instead of the full document:
+
+```shell
+docker buildx imagetools inspect ghcr.io/nvidia/openshell/gateway:latest --format '{{ range (index .SBOM "linux/amd64").SPDX.packages }}{{ .name }}@{{ .versionInfo }}{{ println }}{{ end }}'
+```
+
+The same commands work for `ghcr.io/nvidia/openshell/supervisor`.
+
+## Coverage
+
+Every SBOM lists the base-image packages. Release Dev and Release Tag images also list the Rust crates compiled into their OpenShell binary.
+
+<Note>
+OpenShell also publishes minimal SLSA provenance. It records how BuildKit produced the image, including its source revision, build platform, and base-image materials, without the extra build parameters included by full provenance.
+</Note>

tasks/scripts/docker-build-image.sh

@@ -168,14 +168,19 @@ if [[ "${IS_FINAL_IMAGE}" == "1" ]]; then
 	TAG_ARGS=(-t "${IMAGE_NAME}:${IMAGE_TAG}")
 fi
 
+ATTESTATION_ARGS=(--provenance=false)
 OUTPUT_ARGS=()
 if [[ -n "${DOCKER_OUTPUT:-}" ]]; then
 	OUTPUT_ARGS=(--output "${DOCKER_OUTPUT}")
 elif [[ "${IS_FINAL_IMAGE}" == "1" ]]; then
-	if [[ "${DOCKER_PUSH:-}" == "1" ]]; then
-		OUTPUT_ARGS=(--push)
-	elif [[ "${DOCKER_PLATFORM:-}" == *","* ]]; then
-		OUTPUT_ARGS=(--push)
+	if [[ "${DOCKER_PUSH:-}" == "1" || "${DOCKER_PLATFORM:-}" == *","* ]]; then
+		if ce_is_docker; then
+			# Attestations require a registry-backed image index.
+			ATTESTATION_ARGS=(--provenance=mode=min --attest type=sbom)
+			OUTPUT_ARGS=(--output "type=image,push=true,oci-mediatypes=true,oci-artifact=true")
+		else
+			OUTPUT_ARGS=(--push)
+		fi
 	else
 		OUTPUT_ARGS=(--load)
 	fi
@@ -191,7 +196,7 @@ ce_build \
 	-f "${DOCKERFILE}" \
 	--target "${DOCKER_TARGET}" \
 	${TAG_ARGS[@]+"${TAG_ARGS[@]}"} \
-	--provenance=false \
+	${ATTESTATION_ARGS[@]+"${ATTESTATION_ARGS[@]}"} \
 	"$@" \
 	${OUTPUT_ARGS[@]+"${OUTPUT_ARGS[@]}"} \
 	.

tasks/scripts/verify-image-sbom.sh

@@ -0,0 +1,53 @@
+#!/usr/bin/env bash
+
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+set -euo pipefail
+
+# Validate the final image's SBOM attestation and, for auditable builds, require
+# at least one Cargo package discovered from the binary metadata.
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "${SCRIPT_DIR}/container-engine.sh"
+
+usage() {
+  echo "Usage: verify-image-sbom.sh <image-ref> [--require-cargo]" >&2
+}
+
+IMAGE=${1:-}
+REQUIRE_CARGO=${2:-}
+if [[ -z "${IMAGE}" || $# -gt 2 || ( -n "${REQUIRE_CARGO}" && "${REQUIRE_CARGO}" != "--require-cargo" ) ]]; then
+  usage
+  exit 2
+fi
+
+if ! ce_is_docker; then
+  echo "Error: SBOM attestations are produced on the Docker/buildx path; ${CONTAINER_ENGINE} has no imagetools equivalent" >&2
+  exit 2
+fi
+
+echo "==> Inspecting SBOM attestation of ${IMAGE}"
+SBOM_JSON="$(ce buildx imagetools inspect "${IMAGE}" --format '{{ json .SBOM }}')"
+COUNTS="$(
+  jq -r '
+    [.. | objects | .SPDX? | select(type == "object")] as $documents
+    | [
+        ($documents | length),
+        ([$documents[] | .. | strings | select(startswith("pkg:cargo/"))] | length)
+      ]
+    | @tsv
+  ' <<<"${SBOM_JSON}"
+)"
+read -r SPDX_COUNT CARGO_COUNT <<<"${COUNTS}"
+
+if [[ "${SPDX_COUNT}" -eq 0 ]]; then
+  echo "Error: ${IMAGE} carries no SPDX SBOM" >&2
+  exit 1
+fi
+if [[ "${REQUIRE_CARGO}" == "--require-cargo" && "${CARGO_COUNT}" -eq 0 ]]; then
+  echo "Error: ${IMAGE} SBOM contains no Cargo packages" >&2
+  exit 1
+fi
+
+echo "SBOM attestation verified: SPDX=${SPDX_COUNT}, Cargo=${CARGO_COUNT}"