Thanksgiving 2023 security incident

On Thanksgiving Day, November 23, 2023, Cloudflare detected a threat actor on our self-hosted Atlassian server. Our security team immediately began an investigation, cut off the threat actor’s access, and on Sunday, November 26, we brought in CrowdStrike’s Forensic team to perform their own independent analysis.

Yesterday, CrowdStrike completed its investigation, and we are publishing this blog post to talk about the details of this security incident.

We want to emphasize to our customers that no Cloudflare customer data or systems were impacted by this event. Because of our access controls, firewall rules, and use of hard security keys enforced using our own Zero Trust tools, the threat actor’s ability to move laterally was limited. No services were implicated, and no changes were made to our global network systems or configuration. This is the promise of a Zero Trust architecture: it’s like bulkheads in a ship where a compromise in one system is limited from compromising the whole organization.

From November 14 to 17, a threat actor did reconnaissance and then accessed our internal wiki (which uses Atlassian Confluence) and our bug database (Atlassian Jira). On November 20 and 21, we saw additional access indicating they may have come back to test access to ensure they had connectivity.

They then returned on November 22 and established persistent access to our Atlassian server using ScriptRunner for Jira, gained access to our source code management system (which uses Atlassian Bitbucket), and tried, unsuccessfully, to access a console server that had access to the data center that Cloudflare had not yet put into production in São Paulo, Brazil.

They did this by using one access token and three service account credentials that had been taken, and that we failed to rotate, after the Okta compromise of October 2023. All threat actor access and connections were terminated on November 24 and CrowdStrike has confirmed that the last evidence of threat activity was on November 24 at 10:44.

(Throughout this blog post all dates and times are UTC.)

Even though we understand the operational impact of the incident to be extremely limited, we took this incident very seriously because a threat actor had used stolen credentials to get access to our Atlassian server and accessed some documentation and a limited amount of source code. Based on our collaboration with colleagues in the industry and government, we believe that this attack was performed by a nation state attacker with the goal of obtaining persistent and widespread access to Cloudflare’s global network.

On November 24, after the threat actor was removed from our environment, our security team pulled in all the people they needed across the company to investigate the intrusion and ensure that the threat actor had been completely denied access to our systems, and to ensure we understood the full extent of what they accessed or tried to access.

Then, from November 27, we redirected the efforts of a large part of the Cloudflare technical staff (inside and outside the security team) to work on a single project dubbed “Code Red”. The focus was strengthening, validating, and remediating any control in our environment to ensure we are secure against future intrusion and to validate that the threat actor could not gain access to our environment. Additionally, we continued to investigate every system, account and log to make sure the threat actor did not have persistent access and that we fully understood what systems they had touched and which they had attempted to access.

CrowdStrike performed an independent assessment of the scope and extent of the threat actor’s activity, including a search for any evidence that they still persisted in our systems. CrowdStrike’s investigation provided helpful corroboration and support for our investigation, but did not bring to light any activities that we had missed. This blog post outlines in detail everything we and CrowdStrike uncovered about the activity of the threat actor.

The only production systems the threat actor could access using the stolen credentials was our Atlassian environment. Analyzing the wiki pages they accessed, bug database issues, and source code repositories, it appears they were looking for information about the architecture, security, and management of our global network; no doubt with an eye on gaining a deeper foothold. Because of that, we decided a huge effort was needed to further harden our security protocols to prevent the threat actor from being able to get that foothold had we overlooked something from our log files.

Our aim was to prevent the attacker from using the technical information about the operations of our network as a way to get back in. Even though we believed, and later confirmed, the attacker had limited access, we undertook a comprehensive effort to rotate every production credential (more than 5,000 individual credentials), physically segment test and staging systems, performed forensic triages on 4,893 systems, reimaged and rebooted every machine in our global network including all the systems the threat actor accessed and all Atlassian products (Jira, Confluence, and Bitbucket).

The threat actor also attempted to access a console server in our new, and not yet in production, data center in São Paulo. All attempts to gain access were unsuccessful. To ensure these systems are 100% secure, equipment in the Brazil data center was returned to the manufacturers. The manufacturers’ forensic teams examined all of our systems to ensure that no access or persistence was gained. Nothing was found, but we replaced the hardware anyway.

We also looked for software packages that hadn’t been updated, user accounts that might have been created, and unused active employee accounts; we went searching for secrets that might have been left in Jira tickets or source code, examined and deleted all HAR files uploaded to the wiki in case they contained tokens of any sort. Whenever in doubt, we assumed the worst and made changes to ensure anything the threat actor was able to access would no longer be in use and therefore no longer be valuable to them.

Every member of the team was encouraged to point out areas the threat actor might have touched, so we could examine log files and determine the extent of the threat actor’s access. By including such a large number of people across the company, we aimed to leave no stone unturned looking for evidence of access or changes that needed to be made to improve security.

The immediate “Code Red” effort ended on January 5, but work continues across the company around credential management, software hardening, vulnerability management, additional alerting, and more.

The attack started in October with the compromise of Okta, but the threat actor only began targeting our systems using those credentials from the Okta compromise in mid-November.

The following timeline shows the major events:

We’ve written about this before but, in summary, we were (for the second time) the victim of a compromise of Okta’s systems which resulted in a threat actor gaining access to a set of credentials. These credentials were meant to all be rotated.

Unfortunately, we failed to rotate one service token and three service accounts (out of thousands) of credentials that were leaked during the Okta compromise.

One was a Moveworks service token that granted remote access into our Atlassian system. The second credential was a service account used by the SaaS-based Smartsheet application that had administrative access to our Atlassian Jira instance, the third account was a Bitbucket service account which was used to access our source code management system, and the fourth was an AWS environment that had no access to the global network and no customer or sensitive data.

The one service token and three accounts were not rotated because mistakenly it was believed they were unused. This was incorrect and was how the threat actor first got into our systems and gained persistence to our Atlassian products. Note that this was in no way an error on the part of Atlassian, AWS, Moveworks or Smartsheet. These were merely credentials which we failed to rotate.

Our logs show that the threat actor started probing and performing reconnaissance of our systems beginning on November 14, looking for a way to use the credentials and what systems were accessible. They attempted to log into our Okta instance and were denied access. They attempted access to the Cloudflare Dashboard and were denied access.

Additionally, the threat actor accessed an AWS environment that is used to power the Cloudflare Apps marketplace. This environment was segmented with no access to global network or customer data. The service account to access this environment was revoked, and we validated the integrity of the environment.

The threat actor successfully accessed Atlassian Jira and Confluence on November 15 using the Moveworks service token to authenticate through our gateway, and then they used the Smartsheet service account to gain access to the Atlassian suite. The next day they began looking for information about the configuration and management of our global network, and accessed various Jira tickets.

The threat actor searched the wiki for things like remote access, secret, client-secret, openconnect, cloudflared, and token. They accessed 36 Jira tickets (out of a total of 2,059,357 tickets) and 202 wiki pages (out of a total of 194,100 pages).

The threat actor accessed Jira tickets about vulnerability management, secret rotation, MFA bypass, network access, and even our response to the Okta incident itself.

The wiki searches and pages accessed suggest the threat actor was very interested in all aspects of access to our systems: password resets, remote access, configuration, our use of Salt, but they did not target customer data or customer configurations.

The threat actor used the Smartsheet credential to create an Atlassian account that looked like a normal Cloudflare user. They added this user to a number of groups within Atlassian so that they’d have persistent access to the Atlassian environment should the Smartsheet service account be removed.

During this period, the attacker took a break from accessing our systems (apart from apparently briefly testing that they still had access) and returned just before Thanksgiving.

Since the Smartsheet service account had administrative access to Atlassian Jira, the threat actor was able to install the Sliver Adversary Emulation Framework, which is a widely used tool and framework that red teams and attackers use to enable “C2” (command and control), connectivity gaining persistent and stealthy access to a computer on which it is installed. Sliver was installed using the ScriptRunner for Jira plugin.

This allowed them continuous access to the Atlassian server, and they used this to attempt lateral movement. With this access the Threat Actor attempted to gain access to a non-production console server in our São Paulo, Brazil data center due to a non-enforced ACL. The access was denied, and they were not able to access any of the global network.

Over the next day, the threat actor viewed 120 code repositories (out of a total of 11,904 repositories). Of the 120, the threat actor used the Atlassian Bitbucket git archive feature on 76 repositories to download them to the Atlassian server, and even though we were not able to confirm whether or not they had been exfiltrated, we decided to treat them as having been exfiltrated.

The 76 source code repositories were almost all related to how backups work, how the global network is configured and managed, how identity works at Cloudflare, remote access, and our use of Terraform and Kubernetes. A small number of the repositories contained encrypted secrets which were rotated immediately even though they were strongly encrypted themselves.

We focused particularly on these 76 source code repositories to look for embedded secrets, (secrets stored in the code were rotated), vulnerabilities and ways in which an attacker could use them to mount a subsequent attack. This work was done as a priority by engineering teams across the company as part of “Code Red”.

As a SaaS company, we’ve long believed that our source code itself is not as precious as the source code of software companies that distribute software to end users. In fact, we’ve open sourced a large amount of our source code and speak openly through our blog about algorithms and techniques we use. So our focus was not on someone having access to the source code, but whether that source code contained embedded secrets (such as a key or token) and vulnerabilities.

Our security team was alerted to the threat actor’s presence at 16:00 and deactivated the Smartsheet service account 35 minutes later. 48 minutes later the user account created by the threat actor was found and deactivated. Here’s the detailed timeline for the major actions taken to block the threat actor once the first alert was raised.

15:58 - The threat actor adds the Smartsheet service account to an administrator group.16:00 - Automated alert about the change at 15:58 to our security team.16:12 - Cloudflare SOC starts investigating the alert.16:35 - Smartsheet service account deactivated by Cloudflare SOC.17:23 - The threat actor-created Atlassian user account is found and deactivated.17:43 - Internal Cloudflare incident declared.21:31 - Firewall rules put in place to block the threat actor’s known IP addresses.

10:44 - Last known threat actor activity.11:59 - Sliver removed.

Throughout this timeline, the threat actor tried to access a myriad of other systems at Cloudflare but failed because of our access controls, firewall rules, and use of hard security keys enforced using our own Zero Trust tools.

To be clear, we saw no evidence whatsoever that the threat actor got access to our global network, data centers, SSL keys, customer databases or configuration information, Cloudflare Workers deployed by us or customers, AI models, network infrastructure, or any of our datastores like Workers KV, R2 or Quicksilver. Their access was limited to the Atlassian suite and the server on which our Atlassian runs.

A large part of our “Code Red” effort was understanding what the threat actor got access to and what they tried to access. By looking at logging across systems we were able to track attempted access to our internal metrics, network configuration, build system, alerting systems, and release management system. Based on our review, none of their attempts to access these systems were successful. Independently, CrowdStrike performed an assessment of the scope and extent of the threat actor’s activity, which did not bring to light activities that we had missed and concluded that the last evidence of threat activity was on November 24 at 10:44.

We are confident that between our investigation and CrowdStrike’s, we fully understand the threat actor’s actions and that they were limited to the systems on which we saw their activity.

This was a security incident involving a sophisticated actor, likely a nation-state, who operated in a thoughtful and methodical manner. The efforts we have taken ensure that the ongoing impact of the incident was limited and that we are well-prepared to fend off any sophisticated attacks in the future. This required the efforts of a significant number of Cloudflare’s engineering staff, and, for over a month, this was the highest priority at Cloudflare. The entire Cloudflare team worked to ensure that our systems were secure, the threat actor’s access was understood, to remediate immediate priorities (such as mass credential rotation), and to build a plan of long-running work to improve our overall security based on areas for improvement discovered during this process.

We are incredibly grateful to everyone at Cloudflare who responded quickly over the Thanksgiving holiday to conduct an initial analysis and lock out the threat actor, and all those who contributed to this effort. It would be impossible to name everyone involved, but their long hours and dedicated work made it possible to undertake an essential review and change of Cloudflare’s security while keeping our global network running and our customers’ service running.

We are grateful to CrowdStrike for having been available immediately to conduct an independent assessment. Now that their final report is complete, we are confident in our internal analysis and remediation of the intrusion and are making this blog post available.

IOCsBelow are the Indications of Compromise (IOCs) that we saw from this threat actor. We are publishing them so that other organizations, and especially those that may have been impacted by the Okta breach, can search their logs to confirm the same threat actor did not access their systems.