Thanks for filing this. This is an important issue to me and I want to be sure this works - we want an e2e test for this. I haven't yet context switched to do that. It is certainly possible there are bugs here.

This is also related to containers/skopeo#1829

Hi, has there been any progress on this?

Upon initial review, changes might need to be made to containers/container-image-proxy-rs and skopeo/proxy.go to be able to correctly verify the signed images by sigstore. Right now podman pull with the following configurations in /etc/containers/registries.d/ and /etc/containers/policy.json does identify incorrect signature from sigstore and will give an error. Thus those changes made to containers/image may need to be used in `containers/containers-image-proxy-rs'

Deeper inspection revealed that, rpm-ostree rebase only handles authfiles needed to access private registries. It does not actually do any policy verification of the image or the author afterwards. (except checking if /etc/containers/policy.json has insecureAcceptAnything)

• What podman pull does that rpm-ostree rebase does not? :
  Policy verification from container/image/signature

• How does current rpm-ostree rebase work?
  It pulls the container-image from registry using ostree-ext APIs.
  If it instead used the Image() function from containers/image meant for this purpose and also used by podman pull and skopeo copy, it would by default do this verification.
  However currently the logic of pulling an ostree-based container image and unencapsulating it into an ostree repository are intertwined in this function ImageImporter::Import()

• What are possible steps forward?

  1. Separate the logic of pull and unencapsulate of a container image. Then we can simply call skopeo copy and get pull and verify in one step. Then unencaspulate in the next step.
  2. Add the logic of policy_verification in containers/containers-image-proxy-rs. Basically having a rust binding for the go function. Then invoke the policy verification while doing ImageImporter::Prepare()

@cgwalters any preferences in the solution forward?

Separate the logic of pull and unencapsulate of a container image. Then we can simply call skopeo copy and get pull and verify in one step. Then unencaspulate in the next step.

In that model we can't reuse already fetched layers, forcing every update to be an entire re-pull; it's a nonstarter.

Add the logic of policy_verification in containers/containers-image-proxy-rs. Basically having a rust binding for the go function. Then invoke the policy verification while doing ImageImporter::Prepare()

This SGTM!

Did you see the discussion in containers/skopeo#1829 btw? That's very related to the latter path.

Did you see the discussion in containers/skopeo#1829 btw? That's very related to the latter path.

yes also looking over containers/skopeo#1844. Insightful

I stumbled on this today, and had a look at a few PRs. I am not sure what's is blocking this. Is there a way I can help ? I'm happy to jump in and learn the codebase to contribute the missing bits if I can get a little guidance :)

Fixed by containers/skopeo#2029
A bump to container-image-proxy-rs should introduce the changes in rpm-ostree

Let's keep this open until we have a bit more verification of the fix.

A bump to container-image-proxy-rs should introduce the changes in rpm-ostree

I think actually it's just bumping the skopeo package right?

Hopefully this will help with verification!

https://universal-blue.org/blog/2023/07/19/sigstore-signed-images-are-now-here/

Thanks everyone for working on this!

This does not appear to work for containers signed via the "keyless" method in github action workflows. A github workflow to sign a container using the github actions OIDC token as detailed in the cosign-installer examples results in a container signed with a certificate that doesn't have an email address as the SAN. The SAN is instead a URI to the github workflow yaml file.

See this example workflow with cosign sign steps:

name: Build and push ostree containers
on:
  workflow_dispatch:
  push:
    branches:
      - 'main'
      - 'test'
      - 'stage'
      - 'next'
jobs:
  build-push-base-image:
    name: Build and push baseline workstation image
    runs-on: ubuntu-latest
    timeout-minutes: 60
    permissions:
      contents: read
      id-token: write
      packages: write
    steps:
      - name: checkout repo
        uses: actions/checkout@v3
      - name: cosign-installer
        uses: sigstore/[email protected]
      - name: build base image
        id: build-base-image
        uses: redhat-actions/buildah-build@v2
        with:
          image: fedora-silverblue-ws
          tags: 38-${{ github.ref_name }} ${{ github.sha }}
          containerfiles: ./Containerfile.workstation
      - name: push base image
        id: push-base-image
        uses: redhat-actions/push-to-registry@v2
        with:
          image: ${{ steps.build-base-image.outputs.image }}
          tags: ${{ steps.build-base-image.outputs.tags }}
          registry: ghcr.io/${{ github.repository_owner }}
          username: ${{ github.repository_owner }}
          password: ${{ secrets.GITHUB_TOKEN }}
      - name: cosign login (rh push-to-registry does not login)
        run: cosign login ghcr.io -u ${{ github.repository_owner }} -p ${{ secrets.GITHUB_TOKEN }}
      - name: sign base container image with sigstore fulcio and github oidc
        run: cosign sign --yes ghcr.io/${{ github.repository_owner }}/${{ steps.build-base-image.outputs.image }}@${{ steps.push-base-image.outputs.digest }}

Now, extract the certificate from the signature in the registry:

$ skopeo inspect docker://ghcr.io/jmpolom/fedora-silverblue-ws@sha256:48bb3ee03f4687f86ac0c4d095e096dfa547003321f2519f2d547a2a7dd81ab0 | jq -r .LayersData[0].Annotations.\"dev.sigstore.cosign/certificate\" > cert

And dump the text of the cert:

$ openssl x509 -text -noout -in cert
Certificate:
    (...)
        X509v3 extensions:
            (...)
            X509v3 Subject Alternative Name: critical
                URI:https://github.com/jmpolom/fedora-ostree-ws-container/.github/workflows/build.yml@refs/heads/main

The SAN is a URI and not an email address but containers-policy.json only allows a subject email for the certificate identity. Sigstore seems to be moving away from only allowing email addresses as identities. There are some relevant discussions in issues within their projects. I do not believe this is a sigstore/cosign/fulcio issue whatsoever.

The resulting signed container is perfectly verifiable using cosign verify but podman nor rpm-ostree seem capable of verifying it given what can be configured with container policies. The only reference for subject identity in the containers-policy.json man page is subjectEmail and we do not have a certificate with an email address in it here.

Here is my entire policy.json file:

{
  "default": [
    {
      "type": "reject"
    }
  ],
  "transports": {
    "docker": {
      "registry.access.redhat.com": [
        {
          "type": "signedBy",
          "keyType": "GPGKeys",
          "keyPath": "/etc/pki/rpm-gpg/RPM-GPG-KEY-redhat-release"
        }
      ],
      "registry.redhat.io": [
        {
          "type": "signedBy",
          "keyType": "GPGKeys",
          "keyPath": "/etc/pki/rpm-gpg/RPM-GPG-KEY-redhat-release"
        }
      ],
      "ghcr.io/jmpolom": [
        {
          "type": "sigstoreSigned",
          "fulcio": {
            "caPath": "/etc/pki/containers/fulcio.sigstore.dev.pub",
            "oidcIssuer": "https://token.actions.githubusercontent.com",
            "subjectEmail": "[email protected]"
          },
          "rekorPublicKeyPath": "/etc/pki/containers/rekor.sigstore.dev.pub",
          "signedIdentity": {
            "type": "matchRepository"
          }
        }
      ],
      "": [
        {
          "type": "insecureAcceptAnything"
        }
      ]
    },
    "docker-daemon": {
      "": [
        {
          "type": "insecureAcceptAnything"
        }
      ]
    }
  }
}

For kicks I did try putting the URI from the certificate SAN into the subjectEmail field in policy.json but of course it did not work. Is there a secret configuration key to use here for non-Email subject identities? I assume this simply isn't supported yet within skopeo/podman/rpm-ostree.

@cgwalters have you had a chance to review this yet? Presumably what I'm attempting to do is beyond the capabilities of rpm-ostree/skopeo at this time.

Bump on this again for @cgwalters and @RishabhSaini

My conclusion at this point is the sigstore/cosign verification support available in rpm-ostree, podman and skopeo cannot support the keyless cosign signature workflow that works within CI workflows. I think the way containers policy.json allows specifying the certificate subject requires generalization to permit specifying a URI in addition to email.

Is there any chance that this change might get picked up in rpm-ostree? It is nice to see support for cosign signed containers however the support appears incomplete.

Hi, sorry for the delayed reply. We are going to look at doing some work on this issue soon I hope; pulling it more closely into the priority list.

It is possible to verify keyless signatures with podman - let me try locally to see if I can find the problem with rpm-ostree - and if there is an issue let's get it fixed because it should be working as expected.

It is possible to verify keyless signatures with podman

So I don't disagree that podman can verify keyless signatures in the general sense. The problem is keyless signatures made in CI via github actions (in this specific case, I have not tried gitlab CI). Could you share if you're referring to verifying keyless signatures with podman created in a CI job? The certificate SAN is different and none of the documentation I have reviewed for containers policy.json indicates how I can specify a certificate identity other than an email address (Fulcio certificates for CI jobs do not have an email address identity).

Could you share if you're referring to verifying keyless signatures with podman created in a CI job?

I see now - I read too quickly - if this is a limitation it should not be, I'm going to dig in to fix this (but first I'll be sure it's not just a docs limitation)

Is rpm-ostree using the signature verification logic in containers/image and containers/skopeo? I do see some skopeo proxy errors when signature validation fails. I very quickly looked through the skopeo and image repos and only found references to subjectEmail for Fulcio certs.

If you need any package builds tested let me know, I would be able to test.

@jmpolom I used a github workflow to sign a container using the github actions OIDC token, just as you have done in a previous message of yours. However, my cert still uses an email address as the SAN and not a URI, so I was not able to reproduce the container image.

containers/skopeo#2182 should fix your problem since I've been able to configure my policy.json and rebase to your signed container image. If you would like to test it, you will just need to compile the skopeo binary from source based on my fork and replace it with the one found at /usr/bin/skopeo on your system.

I used a github workflow to sign a container using the github actions OIDC token, just as you have done in a previous message of yours. However, my cert still uses an email address as the SAN and not a URI, so I was not able to reproduce the container image.

Could you link me to the exact github workflow you used? I would like to review it. I was trying to see if there was a way to switch to an email address when performing the signature in a workflow using the OIDC token but could not find anything. Perhaps it has been added or there has been an upstream change I'm not aware of.

PR looks reasonable though. Figured it would be fairly straightforward to support the URI based SAN. This definitely should be supported functionality.

Question before I can test: on an ostree based system how do I replace something in /usr/bin as that area is immutable by design? I have had bad results attempting to do "live" filesystem edits with rpm-ostree. Is the correct way rpm-ostree override replace?

Could you link me to the exact github workflow you used?

https://github.com/lukewarmtemp/custom-container-images/blob/main/.github/workflows/docker-image.yml

how do I replace something in /usr/bin as that area is immutable by design?

I just use the following to remove immutability and replace the bin in /usr/bin

$ sudo ostree admin unlock

Ah, I think I see why you did not end up with a container signed with a URI as the SAN. Are you defining a github actions secret called GH_PAT?

I believe you are inadvertently using the "workaround" method for the issue described here. It's not really a workaround because you need to enter the one time code from the workflow in a browser manually so your builds are not really automatic at that point.

What you want to use is the GITHUB_TOKEN secret for the workflow job. This is also somewhat indirectly mentioned in the cosign-installer actions docs in the example where they do a docker login. The github actions workflow that I am using uses this ephemeral workflow token and is what results in the URI in the certificate SAN.

I will try to test this later this week. You may want to revise your test workflow to use the workflow token instead of a user defined access token.

I will try to test this later this week. You may want to revise your test workflow to use the workflow token instead of a user defined access token.

Hey @jmpolom @sallyom before trying this again I may ask for your results? 😅

Hey @jmpolom @sallyom before trying this again I may ask for your results? 😅

This occurred so far in the past I no longer recall what I was going to test. I don't think I tested anything specific because I probably determined there was in fact nothing to test. I have a container image regularly being signed and I'll probably be building a system this week with it I'll try to pull with signature verification enabled.

The PR by @lukewarmtemp did not appear to add the necessary functionality (logic seemed backwards) to resolve this issue. The PR is is also seemingly stalled on ideological grounds by the maintainer @mtrmac not wanting to add support for pulling the identity from the certificate SAN. See the PR for the details. Maybe someone else asking for this to "just work" (R) as it should might get things moving in the right direction.

This remains an open issue and this functionality does not work as needed, as of today.

@jmpolom @sallyom @frzifus @cgwalters

Is rpm-ostree using the signature verification logic in containers/image and containers/skopeo? .... I very quickly looked through the skopeo and image repos and only found references to subjectEmail for Fulcio certs.

Like @jmpolom mentioned above, the issue is not with rpm-ostree but with github.com/containers/image - it only accepts fulcio.subjectEmail identity - with GH actions and/or workload identity, there is no email, instead the signature/cert references the GH action workflow, something like this

Issuer: https://token.actions.githubusercontent.com
Subject: https://github.com/sallyom/image-build/.github/workflows/build-sign-push.yaml@refs/heads/main

I verified that when signing with the non-gh-action manual browser flow where signature has:

"Issuer":"https://github.com/login/oauth","Subject":"[email protected]"}}

The NON-bootc (NON-rpm-ostree) images are properly gated with a carefully configured containers/policy.json.

@mtrmac will be helpful here - looks like containers/image needs to learn about gh-actions/workload-identity

However, I cannot pull a bootable container image with the same configuration. So there seems to be 2 issues here:

1. containers/image doesn't know about workload-identities, only email identities
2. not sure why but containers/image doesn't recognize a signature - even an email-identity sig - attached to a bootable image.

so I don't have to look this up again to test, here's the /etc/containers/policy.json that works (with emailSubject only):

    "transports": {
        "docker": {
            "quay.io/sallyom": [
                {
                    "type":    "sigstoreSigned",
                    "fulcio": {
                        "caPath": "/etc/.sigstore/fulcio_v1.crt.pem",
                        "oidcIssuer": "https://github.com/login/oauth",
                        "subjectEmail": "[email protected]"
                    },
                    "rekorPublicKeyPath": "/etc/.sigstore/rekor.pub",
                    "signedIdentity": { "type": "matchRepository" }
                }
            ],

but what we want to work is this for fulcio (really need only the added subject, the Issuer is not the problem):

                    "fulcio": {
                        "caPath": "/etc/.sigstore/fulcio_v1.crt.pem",
                        "oidcIssuer": "https://token.actions.githubusercontent.com"",
                        "subject": "https://github.com/sallyom/image-build/.github/workflows/build-sign-push.yaml@refs/heads/main"

and also need this in /etc/containers/registries.d/default.yaml - but this hurts, i'll open an issue in containers/image to use-sigstore-attachments by default

docker:
  quay.io/sallyom:
    use-sigstore-attachments: true

but what we want to work is this for fulcio:

                    "fulcio": {
                        "caPath": "/etc/.sigstore/fulcio_v1.crt.pem",
                        "oidcIssuer": "https://token.actions.githubusercontent.com"",
                        "subject": "https://github.com/sallyom/image-build/.github/workflows/build-sign-push.yaml@refs/heads/main"

@sallyom It's actually a URI SAN, so you probably want subjectURI key (see containers/image#2235 (comment) and refer to the RFC). Or generalize the subject to match arbitrary text for what appears in the certificate SAN? Either should work if the maintainer agrees with it.

The NON-bootc (NON-rpm-ostree) images are properly gated with a carefully configured containers/policy.json.

Is keyless sigstore signature verification completely broken for ostree containers? That would seem to be a worse issue than depicted here. I don't think I ever really could get this working so I can't recall if I ever got a signed image pulled via rpm-ostree or such tools.

@mtrmac will be helpful here - looks like containers/image needs to learn about gh-actions/workload-identity

I have been back and forth with @mtrmac about this in the draft PR I linked above. He does not seem interested in adding support for matching other texts (like a URI) in the certificate SAN. He believes this would somehow lead to users configuring insecure policies. I think the risk is low and his rationale seemed to be second guessing the decision made by the sigstore community to place certificate identity into the SAN. My thought is using anything else besides the SAN risks compatibility issues down the road?

I was considering trying to add in the needed support but I'm not even clear it would be merged due the strong opinions of the maintainer. A fix for this issue has basically been stalled out because of this. There really wasn't a counter proposal from him as to what he'd deem acceptable to merge.

However, I cannot pull a bootable container image with the same configuration. So there seems to be 2 issues here:

containers/image doesn't know about workload-identities, only email identities
not sure why but containers/image doesn't recognize a signature - even an email-identity sig - attached to a bootable image.

The NON-bootc (NON-rpm-ostree) images are properly gated with a carefully configured containers/policy.json.

Is keyless sigstore signature verification completely broken for ostree containers? That would seem to be a worse issue than depicted here. I don't think I ever really could get this working so I can't recall if I ever got a signed image pulled via rpm-ostree or such tools.

Is keyless sigstore signature verification completely broken for ostree containers? That would seem to be a worse issue than depicted here. I don't think I ever really could get this working so I can't recall if I ever got a signed image pulled via rpm-ostree or such tools.

I am hoping this is not the case - one thing about testing yesterday is I was using multi-arch manifest list - maybe that was the issue? I will test more today - I'm testing with bootc images instead of rpm-ostree but am guessing whatever we find applies to both: https://docs.fedoraproject.org/en-US/bootc/rpm-ostree/

With bluebuild we've been trying to find ways to use OIDC. Having the ability to use keyless verification would be perfect for our user base. They wouldn't have the know how for managing keys. I'd argue not having this creates friction for less technical users.

I am hoping this is not the case - one thing about testing yesterday is I was using multi-arch manifest list - maybe that was the issue? I will test more today - I'm testing with bootc images instead of rpm-ostree but am guessing whatever we find applies to both: https://docs.fedoraproject.org/en-US/bootc/rpm-ostree/

I have a single architecture derivative I'm building that gets signed keylessly via the github OIDC token (so URI SAN). Please feel free to use it as a test case: https://github.com/jmpolom/fedora-ostree-ws-container/pkgs/container/fedora-silverblue-ws

As far as I've seen all the centos/fedora bootc images at the end of the day still get built with rpm-ostree so there should be no difference. The underlying ostree libraries being used to pull the ostree container image should all be common anyway.