Page tree
Skip to end of metadata
Go to start of metadata

This is mostly a tour of how Jira does CSRF (cross-site request forgery), written for my own benefit while investigating the problems eventually logged as JRASERVER-73880 - Getting issue details... STATUS .


Have you ever tried to attach a file or image to a Jira issue, and got the error Jira could not attach the file as there was a missing token. Please try attaching the file again. ? This commonly happens when:

  1. You are viewing a Jira issue
  2. You leave your computer for over 5 hours --or--  in another tab you log out and back in to Jira.
  3. You return to the issue, and without refreshing the page  try to attach a file or image to the issue.

    It also happens if you click the 'Create' button to create a new issue, and on the 'Attachment' field try to upload an attachment.

Background: CSRF attacks

The 'missing token' in question is a CSRF token used to prevent CSRF (cross site request forgery) attacks.

An example of a CSRF attack would be as follows. An attacker creates a HTML page at https://hax0rs.com containing:

<a href="https://jira.yourcompany.com/secure/admin/user/AddUser.jspa?email=evil%40hax0rs.com&fullname=Mr+Hax0r&username=hacker&password=hacker123&selectedApplications=jira-software&Create=Create+user">Click me!</a>

If you can trick a Jira administrator into visiting that site and clicking that link, a new hacker  account is created in Jira.

 Why did Jira, upon receiving that AddUser.jspa HTTP request, trust the sender? Because your browser identified the request as "coming from you" by including a identifying JSESSIONID cookie, which is scoped to jira.yourcompany.com.

SameSite=Lax

Until 2020, you could even embed the malicious URL in a <img>  tag, and not even need the user to click That is no longer possible since SameSite=Lax became the default, 'Lax' meaning the JSESSIONID  cookie is sent when a link is clicked, but not for non-interactive URLs like in <img>.

Jira's CSRF defenses

The atl_token CSRF token

Jira's main defense against CSRF is to generate a random bit of text (the 'CSRF token') just for you when you log in, and embed it in every HTML form that, if submitted, would changes server state (like adding users). The field is called atl_token:

When the user submits a form in Jira (on the 'Add User' admin page, for example), this atl_token parameter is submitted in the POST body. Jira checks that the atl_token  value is correct before performing the operation. The CSRF token value is a secret known only to Jira and your browser, and changes after every login or session expiry (5h of inactivity). An attacker won't know the correct atl_token  parameter to include in their forged URL.

This is known in the literature as the synchronizer token pattern (OWASP)

Only some Jira operations get token-based CSRF protection. It seems a bit haphazard, e.g. creating a Jira issue is CSRF protected, but adding a comment isn't. Specifically, CSRF protection is applied to:

  • *.jspa Webwork actions like CreateIssueDetails.jspa, whose execute()  method has a @requiresXsrfCheck  annotation. For the implementation, see JiraActionFactory.java which invokes DefaultXsrfInvocationChecker.java.
  • POST, DELETE AND put operations on REST resources, like , submitted via HTTP forms. See Atlassian REST API design guidelines for the rules here. This is implemented in XsrfResourceFilter, which also eventually leads to DefaultXsrfInvocationChecker.java  

Origin/Referer checking

In addition to the atl_token mechanism, Jira also implements the 'Defence In Depth Techniques' on the OWASP page. Specifically, the Origin  and Referer  request headers are checked for REST resources (but perhaps not yet for .jspa  requests - see JRASERVER-63915 - Getting issue details... STATUS ). Origin  and Referer  tell Jira where the request originated – https://hax0rs.com in our example.  This is implemented in XsrfResourceFilter.java .


Jira's CSRF Implementation Details

Back to those atl_token  parameters.

We know that the token is initially generated by Jira when you log in, and is embedded in the returned HTML:

But where does Jira store its copy of the parameter? In two places actually:

  1. the user session
  2. in the user's browser as a cookie.

The atlassian.xsrf.token  session attribute

While you are actively using Jira, Jira maintains a 'session' for you. A session is just a handful of key:value pairs in memory, grouped by 'session id', which is associated with your browser requests with the JSESSIONID cookie your browser sends.

If you are a Jira administator you can see your session's attributes by dropping sessionattributes.jsp into your atlassian-jira/secure/  directory, and hitting the /secure/sessionattribute.jsp  URL:

There in the session, you see where Jira is storing your CSRF token: in the atlassian.xsrf.token attribute (Jira uses 'XSRF' as a synonym for CSRF).

There are a few more wrinkles, but that is the essence of Jira's CSRF protection: all state-modifying HTTP requests must have an atl_token field, whose value must match the atlassian.xsrf.token session parameter. 

The atlassian.xsrf.token cookie

If you examine the cookies your browser associates with Jira, you'll see an atlassian.xsrf.token  cookie:

This is sent along with every HTTP request.

It would therefore be easy to think that this atlassian.xsrf.token cookie is the thing in the request that needs validating. Isn't the server just comparing the atlassian.xsrf.token request cookie with the atlassian.xsrf.token session attribute is knows is correct, and if they match, passing the CSRF check?

No! Remember that the atlassian.xsrf.token cookie is sent by the browser, and will always be present in malicious and non-malicious requests. Its value will always be correct, because cookies aren't under the attacker's control. The presence of a correct atlassian.xsrf.token  cookie tells Jira nothing about whether the request has a malicious origin. It is the atl_token  form parameter which needs validating, not the atlassian.xsrf.token request cookie.

In fact, when the server receives a request, it could compare the atl_token  form parameter with either the atlassian.xsrf.token request cookie, or the atlassian.xsrf.token session attribute. In fact Jira does both (in XsrfTokenStrategy.java).

Why does Jira even set the atlassian.xsrf.token cookie? No idea. The general question of "why send the CSRF token as a cookie" is asked and answered on Stackoverflow. I'm not sure if any of those advantages actually apply to Jira. Certainly Jira does not populate the atl_token form field from the cookie, as it might do. and does not implement the "double submit cookie" pattern where the cookie value is compared with the form value, removing the need for server-side state. Perhaps it is handy to have AJAX Javascript-initiated requests automatically CSRF-protected.

When sessions expire..

One nice thing Jira could  have done with the incoming atlassian.xsrf.token  cookie it is re-initialize the atlassian.xsrf.token  session parameter from the cookie if the session has expired. Then if your session expired overnight, and you had a Jira form half-filled in, the submit (next morning) would work, as the session atlassian.xsrf.token would be seeded from the atlassian.xsrf.token  cookie. Currently that doesn't happen: if the session has expired, XsrfTokenAdditionRequestFilter.java (very early in the filter chain) just generates a new atlassian.xsrf.token session attribute, which is then compared to the atl_token field, fails to match, and you get a 'missing token' error to brighten your Monday morning:

..to Be Continued

Apologies to anyone reading, but there is no conclusion yet. The fix suggested above (copy the token from cookie to session) could be implemented as a custom servlet filter, even implemented as a plugin.  If I ever do so I will update this post. 

  • No labels