| title |
|---|
GitLab bot security |
Make sure you understand GitLab's security model, before you run a "bot" service like Renovate on GitLab, particularly the pipeline credentials.
!!! warning If you have any doubts or concerns about this content that could affect other users, please follow our Security Policy and report them confidentially.
The concept of CI_JOB_TOKEN permissions was overhauled in GitLab release 8.12, jobs now run with the permissions of the user account which triggered the pipeline.
For security reasons the token was limited to read-only permissions and a limited set of API endpoints, but it’s been extended to allow write access to the GitLab Package Registry.
Any pipeline triggered by a user account thus has permissions to:
- read any repository which that account has access to
- publish packages to them
With the current GitLab CI permissions model, you should only commit to a project which you trust completely. Because that project could maliciously steal repository data, publish fake releases, or spam releases.
With GitLab's current security model, we find the risks of running a public bot service like Renovate are too high. Therefore we stopped hosting Renovate on GitLab, and are waiting for a better security model.
You should remember that when accounts are invited into projects or groups on GitLab, acceptance happens automatically. This was a useful feature to leverage for a shared service.
If you are running a self-hosted Renovate service, we recommend you:
- Run a shared service only within projects which have shared visibility/security within the users, or which have a low risk that a user would try to gain access to a private project they don't otherwise have access to
- If running with
autodiscover, also configure a value forautodiscoverFilterso that the bot can't be invited to projects or groups you don't intend
The following research notes may help you to assess the GitLab bot security risk.
If you only run a bot service on public projects, the risk of unauthorized users accessing private project data is zero. But malicious users can still spoof or spam packages to any other public project they are not a member of, this rules out this approach for a public hosted service.
A public-visibility-only bot service should be low risk for most self-hosted GitLab instances. But you can't stop users from inviting the bot into private projects by accident, which is risky.
Project Access Tokens (PATs) are a recently added feature for GitLab. The main downsides to using PATs for a shared bot service are:
- You can not provision PATs through the API, so project maintainers would need to provision a project bot account and then save it to Renovate manually and per-project
- PATs are a paid-only feature for gitlab.com, which prevents users on the free plan from using them
- At the time of writing, there are still some issues with getting PATs to trigger and authenticate CI
- Any service using PATs would get MRs from a user like
@project_123_botinstead of@renovate-bot
The big benefit of PATs is their limited scope: users with write access to one project cannot read/write to other projects.
Group Access Tokens are still in the planning stage, but may offer a more scalable way to manage a Renovate service. Tokens could be provisioned into Renovate per-group. Permissions and visibility must be kept uniform throughout the group to prevent a privilege escalation.
Many GitLab users do not have uniform permissions and visibility throughout groups today, so this is a risk of Group Access Tokens in general.
The gitlab-org organization on GitLab shows how common it is to mix project visibility within a same group.
And the same as with PATs, if Group Access Tokens becomes a paid feature then users on a free plan can't use the feature.
The security problem described above is that if a user triggers a malicious pipeline then they can be exploited, so skipping CI altogether would seem to be a way to avoid that. If Renovate can reliably force CI skipping for both (a) branch push, and (b) MR creation/updating then there should be no security exploit risk, but of course then there are no tests run instead. A possibility in future could be to combine this with a force push from a user or project token to trigger tests.
The above solution/workaround will be actively researched in collaboration with GitLab.
An alternative to a bot service running with a bot PAT would be to have it run using user OAuth tokens. In this scenario, an OAuth app would be needed to allow users to "install" the bot into projects with members they trust not to exploit them, and then commits and Merge Requests would appear to be authored by the user, not any bot. Bot services are better if they are provisioned with a "bot identity" so that users can quickly distinguish bot activity from real user activity.
Until we can safely reactivate the hosted app, we recommend users migrate to use self-hosted pipelines to run Renovate. Read the renovate-bot/renovate-runner README on GitLab to learn how.
We're trying to find a workable design for the GitLab app, so we can enable it safely again. If you have any ideas, please open a discussion and let us know!
GitLab introduced Group Access Tokens & API for paid & self-hosted instances, but a good permission setup/flow is still not possible. Check out GitLab issue #346298.
Thank you to Nejc Habjan for bringing this security challenge to our attention, and also to his colleagues at Siemens for their help researching the risks. Thanks also to the GitLab security team for being responsive to our questions.