On May 2, 2016, we released a major security update, primarily to fix a critical security issue that allowed a user to gain administrative access via the "impersonate" feature. Now that some time has passed and most of our users have had sufficient time to upgrade, we'd like to reflect on what happened, how it occurred, and what we're doing in the future to improve security in the GitLab code base.
Since May 2, we have released a number of security updates to address certain vulnerabilities, but none of the updates have addressed a bug as serious as the one in the "impersonate user" feature, which is now known as CVE-2016-4340. We released this feature in November of 2015 in GitLab 8.2. It enables admins to diagnose issues with their GitLab installation by allowing them to see what their users see. Since the initial implementation, members from the community helped improve the code, but nobody noticed there was a security hole in one of the controller methods.
Discovering the hole
All that changed on Friday, April 22, 2016, when Douwe Maan, our Backend Lead, began reviewing the code for the feature. Having reviewed hundreds of merge requests and handled an influx of HackerOne security reports, Douwe immediately spotted something wrong: a critical authorization flaw in one of the API endpoints. This flaw would allow a user to gain full GitLab administrative access. Within an hour, Douwe submitted a fix internally. Over the weekend, we began strategizing about how best to roll out this fix to the community.
Our first question: How do we protect GitLab.com against this vulnerability without disclosing details that might harm users who have their own installations of GitLab? We first considered applying a hotfix to GitLab.com, but our infrastructure team had no desire to do this. Applying hotfixes to a live-running, production system is fraught with perils, and we did not want to risk causing other issues. Instead, our Infrastructure Lead, Pablo Carranza proposed blocking the vulnerable route via a HAProxy rule. This would only be a simple configuration change in one place. The following Tuesday, Pablo applied the HAProxy rule and verified that it successfully blocked the route.
With GitLab.com patched with this workaround, we next had to consider: how much advance notice should we give to our users about a security release? Since GitLab is open source, releasing an update means the code would be available for any malicious user to study how to exploit the hole. At first, we decided on a 3-hour timeline:
- Send a security announcement on our security mailing list (T-3 hours)
- Make the GitLab packages online and update GitLab.com (T-1 hour)
- Announce on the blog (T)
For most zero-day vulnerabilities, vendors simply announce updated packages and encourage users to update immediately. However, after further discussion, we felt releasing GitLab in a 3-hour window would not be responsible due to a number of reasons:
- Most administrators would be caught off-guard without at least 24-hour notice
- The severity and ease-of-exploit of the bug could cause significant problems
- Native package maintainers of GitLab (e.g. Debian, FreeBSD, etc.) would not have updates in time
These reasons convinced us to take the unusual step of giving a notice of an impending release for the following Monday. On Wednesday, April 27 around 5:30 pm UTC, we sent out this announcement to our security mailing list:
We have discovered a critical security issue in all GitLab CE and EE versions from 8.2 to 8.7.
On Monday May 2, 2016 at 5:00pm PDT, we will publish new GitLab patch releases for all affected versions. We strongly recommend that all installations running a version mentioned above be upgraded as soon as possible after the release. Please forward this alert to the appropriate person at your organization and have them subscribe to Security Notices. The following versions are affected:
8.6.0 through 8.6.7
8.5.0 through 8.5.11
8.4.0 through 8.4.9
8.3.0 through 8.3.8
8.2.0 through 8.2.4
In addition, we shared patches privately with GitLab package maintainers, who all appreciated the advanced warning. They immediately began work on updating their native packages with the changes.
The next day, Thursday, we completed the long task of incorporating the patches and building a total of 42 Omnibus packages for all six releases to a private repository. We updated GitLab.com to 8.7.1, which contained the fix, and removed the HAProxy workaround.
Some recipients of the e-mail expressed confusion on Hacker News because they did not see a security announcement on our blog. They suspected the e-mail was spam.
After reading the Hacker News post, CEO Sid Sijbrandij pointed out that announcing the affected versions dramatically reduced the search scope of the bug. An attacker could see what changed between 8.1 and 8.2 and discover the vulnerability. If someone exploited the bug over this weekend, customers would have no way to patch their installations for several days. A discussion ensued about whether to move up the release earlier. Instead of Monday, what about Thursday or Friday? We nixed Thursday because the day was over for our European team; more time was needed to have the packages ready. We considered moving the release up to Friday, but a number of people on the team argued that this was not a good idea. Senior Developer Robert Speicher articulated it well:
We chose the date we did to give people time to prepare. A smart, nefarious person might figure out the exploit over the weekend, but releasing early will catch people off-guard and put the exploit into the wild.
We decided to stay with Monday but prepare everything just in case we needed to release early. In addition, we belatedly posted a blog entry to match our security e-mail notice, but this time we omitted the affected versions.
Releasing to the Public
Monday, May 2, 2016 came without incident, and the day of the rollout went smoothly. Around 11:59 UTC, we transferred all 42 Omnibus packages from our private repository to the public one and pushed up new Docker images to Docker Hub. We published the blog post, code, and disclosed all previously-confidential issues to the public on GitLab.com. The security update hit the front page of Hacker News.
What Went Right
A number of things went right:
- There were no reported incidents of anyone exploiting this bug prior to our disclosure.
- We were able to reproduce, fix, and test the security vulnerabilities quickly.
- Even though we are a remotely-distributed team, we were able to effectively communicate and pull together to get the many tasks done for the security update.
What We Are Doing to Improve
At GitLab, we prioritize security issues and try to address them as soon as possible. Since this release, we have learned a number of things that have been put into action:
- We need a better workflow/tools to produce confidential merge requests and packages. #14567
- When we send out security notices via e-mail, we should always have an accompanying blog post.
- In the future, if we send an early security notice, we will omit the version numbers affected to prevent people zeroing in on the vulnerability.
- We need better abstractions for our permission checking #15450
- We need to hire full-time engineers to focus on improving security and to conduct internal audits of our code
- We need to promote our bug bounty program on sites like HackerOne
In general, we received positive responses to this May 2016 security release. Many of our users understood that giving advance notice for a security update made sense. We thank the GitLab community for their patience and understanding. We will continue to be vigilant about security issues within GitLab.
If you have not already, please sign up on for future security notices on this page.
Join us on July 27th for our joint webcast with Yubico
We recognize that security is a growing concern for a number of teams. We're partnering with Yubico again. This time to discuss industry trends and best practices for security. Here's a quick look at what we'll be talking about.
- Key trends that affect the security of your team
- Real-life examples of how both GitLab and Yubico work to improve security
- Advice on industry best practices
If you're interested in learning more or asking security-related questions, please join us on July 27th. Register here.