Skip to content
Zero Day Playbook

The developer collaboration platform playbook for the first 72 hours

  • 0 to 4 hours: 4 steps
  • 4 to 24 hours: 3 steps
  • 24 to 72 hours: 2 steps
  • Hours elapsed: the bar under the cells

Checklist to print

  1. 0 to 4 hours4 steps, due by hour 4
    1. 1

      Security lead, hand-off

      List every instance of the eight products and read each version from the administration System information page. Match it to the fixed builds: Bitbucket Data Center 9.4.26, 10.2.8 or 10.5.1; Confluence Data Center 9.2.26 or 10.2.19; Jira Service Management Data Center 5.12.40, 10.3.26 or 11.3.12; Jira Software Data Center 9.12.40, 10.3.26 or 11.3.12; Bamboo Data Center 10.2.24 or 12.1.12; Crowd Data Center 6.3.7, 7.0.3, 7.1.7 or 7.2.4; Crucible 4.9.15; Fisheye 4.9.15. Take the fixed build on your own line. If your line is not listed, it is end of life: Atlassian publishes no per-line mapping and says to patch to "the fixed LTS version or later", so move to the oldest listed build for the product (Bitbucket 9.4.26, Confluence 9.2.26, Jira Service Management 5.12.40, Jira Software 9.12.40, Bamboo 10.2.24, Crowd 6.3.7, Crucible and Fisheye 4.9.15) and read Atlassian's upgrade notes for whether your version can go straight there. Atlassian Cloud sites were patched by Atlassian and need no action.

      Done at

    2. 2

      Platform engineering, hand-off

      Take instances reachable from the internet off it, or restrict them to the corporate network or VPN. Atlassian says this applies even to instances with user authentication. In the reverse proxy or firewall, allow the corporate and VPN ranges and deny the rest. Include Bitbucket mirrors and mirror farm nodes, which serve the same web layer.

      Done at

    3. 3

      Platform engineering, hand-off

      Add Atlassian's Option 1 rule at the WAF or proxy for every product: block any URL that matches (?is).*(?:/|\\|::|%(?:25)*(?:2f|5c)|(?::|%(?:25)*3a){2})(?:\.|%(?:25)*2e){2}(?:/|\\|::|%(?:25)*(?:2f|5c)|(?::|%(?:25)*3a){2}|;|%(?:25)*3b|$).* , which means two dots next to a slash, backslash or double colon in plain, URL-encoded and double-encoded form. Then test it from outside with curl --path-as-is -s -o /dev/null -w '%{http_code}\n' against your own host using /x/%2e%2e/y, /x/%252e%252e%5cy and /x/..;/y. All three must be blocked. A path like /x/file..txt must still work.

      Done at

    4. 4

      Security lead, hand-off

      Preserve and search the access logs back to 5 Oct 2026. Copy the logs of the proxy and each product, then URL-decode each request line up to two passes and search for two dots directly next to a slash, backslash or double colon. A quick pass on raw logs: grep -Ei '(/|\\|::|%(25)*(2f|5c))(\.|%(25)*2e){2}(/|\\|::|%(25)*(2f|5c)|;|%(25)*3b|[ ?]|$)' access*.log. Then sort the hits by response size and status. A 200 with a body, on a path under WEB-INF, means that file was read. The vendor regex also catches the form watchTowr described, in which a request to the plugin resource path (the segment named download/resources) uses double colons as parent-directory segments and names a file under WEB-INF. Search separately for any request URL that contains both the text WEB-INF and a double colon: grep -Ei 'WEB-INF' access*.log | grep -E '::|%3a%3a' (run it on the decoded log too). Search separately for requests to the Crowd REST paths /crowd/rest/usermanagement/1/ from addresses that are not your own applications, since a stolen application password is used there.

      Done at

  2. 4 to 24 hours3 steps, due by hour 24
    1. 5

      Platform engineering, hand-off

      Upgrade in this order: Crowd first, because the directory is read by the others and its crowd.properties holds an application password; then the products with internet-reachable instances; then the rest. Back up, shut down each node, upgrade to the fixed build, restart, and confirm the version on the System information page. Do not skip the mirrors: upgrade every Bitbucket mirror and mirror farm node.

      Done at

    2. 6

      Platform engineering, hand-off

      If an instance cannot be upgraded within 24 hours, apply a rewrite rule at the application. Confluence, Jira Service Management, Jira Software, Bamboo and Crowd (Option 2): back up, shut the node down, add <Valve className="org.apache.catalina.valves.rewrite.RewriteValve" /> inside the application's Context element in conf/server.xml (for Crowd, apache-tomcat/conf/Catalina/localhost/crowd.xml or apache-tomcat/conf/server.xml, whichever holds the Context), then put Atlassian's rewrite.config in confluence/WEB-INF, atlassian-jira/WEB-INF, bamboo/WEB-INF or atlassian-bamboo/WEB-INF, or crowd-webapp/WEB-INF, and restart. Each rule pairs a RewriteCond using the regex against %{REQUEST_PATH} (and a second against %{REQUEST_URI} as a raw fallback) with RewriteRule (?s)^.*$ /?denied [F,L]. Bitbucket (Option 3): back up, add a new first rule to app/WEB-INF/urlrewrite.xml on every node, mirror and mirror farm node, with the regex in its from element, status 404 and target /mvc/error404, then restart.

      Done at

    3. 7

      Security lead, hand-off

      List what a reader could take. On each node list the web root (ls -la confluence/WEB-INF, atlassian-jira/WEB-INF, bamboo/WEB-INF, crowd-webapp/WEB-INF, app/WEB-INF) and open any configuration file for credentials. For Crowd, check crowd.properties. Rotate every credential found in a file under the web root if the log search shows a request for it, and treat it as exposed anyway on any instance that was internet-reachable while unpatched. Files outside the web application root are not covered by this flaw.

      Done at

  3. 24 to 72 hours2 steps, due by hour 72
    1. 8

      Application owner, hand-off

      Test what depends on each product before restoring access: application links between Jira, Confluence and Bitbucket, Bamboo agents and their links to Bitbucket, Crowd-backed logins, webhooks and Marketplace apps. Check that the Option 1, 2 or 3 rules did not block a legitimate URL; the vendor warns any URL containing such a sequence is denied. Then restore internet access only through the intended gateway.

      Done at

    2. 9

      Security lead, hand-off

      Close or escalate the compromise question using the evidence: no hits and no exposed credentials means close with the log range noted; hits that returned content, or a CISA forensic triage requirement you are subject to, mean preserve disk images and logs of the node first, then start incident response. Keep the Option 1 rule in place after the upgrade until your next review.

      Done at

Bitbucket, Confluence, Jira Software, Jira Service Management, Bamboo, Crowd, Crucible and Fisheye are rarely patched together. Different teams own them, they run on different release lines, and Crowd sits behind most of the others as the user directory. A single flaw in the shared web layer hits all of them at once. CVE-2026-21589 is an arbitrary file read: an unauthenticated caller who knows the exact name and path of a file inside the web application root can fetch it. It does not list directories, so the damage depends on what sits under the root, such as WEB-INF configuration and, for Crowd, the application password in crowd.properties.

What makes it hard to patch is the spread. Atlassian says every version before the fixed builds is affected, so an instance on a release line that has no fixed build must move to a listed line. Atlassian also says it cannot confirm whether any given instance was affected, and tells you to check all of them for signs of compromise. Fixes ship as maintenance releases, not binary patches, so the upgrade is a normal upgrade with a normal test cycle. The first job is to stop the reachable path, not to start that cycle.

When no fix exists yet

Two situations leave you without a fix. The first is an end-of-life release line: Atlassian says all versions are affected and recommends the fixed LTS version or later, so a line with no fixed build must be upgraded to a listed line, and until then the WAF rule, the Tomcat rewrite rule or the Bitbucket urlrewrite rule is the only protection. The second is a fixed build you cannot test inside 24 hours. Both cost you something: the rewrite options need a restart on every node and deny any URL with the sequence, and the WAF rule is only as good as your test of the encoded forms.

Take the instance off the internet at once if it is reachable, and keep it off until it is patched or protected by a tested rewrite rule. The flaw gives read access to files with a known path, so the decision turns on what is under the web root: if the root holds credentials you cannot rotate quickly, or Crowd is exposed, take the service down rather than leave it running behind a rule you have not proven. Atlassian itself advises removing instances from the internet until you can patch or apply mitigations.

CVE-2026-21589 worked through, day by day

Atlassian published the advisory on 5 Oct 2026 for Bitbucket, Confluence, Jira Service Management, Jira Software, Bamboo and Crowd Data Center, plus Crucible and Fisheye. CVSS 4.0 base score 9.3. CISA lists the flaw in KEV, but neither Atlassian nor CISA publishes a date of first exploitation. The dated events are the advisory, a research post and the CISA listing, and they are given by day.

WhenWhat happenedSource
Before 5 Oct 2026Atlassian has patched its Cloud products and reports that its investigation found no evidence of exploitation there.source
5 Oct 2026, Day 0Atlassian publishes the advisory with fixed versions for all eight products and three mitigation options. NVD publishes the record the same day.source
6 Oct 2026, Day 1watchTowr Labs publishes its analysis. The cause is the web resource handler in atlassian-plugins-webresource (6.0.7 vulnerable, 6.0.8 changed), which turns a double colon in a request path into a slash. A request to the plugin resource path of Jira, built with double colons as parent-directory segments, returns WEB-INF/web.xml.source
6 Oct 2026, Day 1The same post gives the request forms for the other products: Confluence uses its plugin resource path under the /s/ static prefix, and Bitbucket does the same with its server plugin, naming a different WEB-INF file because web.xml is blocked there. The researchers read files in the web application root only; they could not leave the Tomcat context.source
6 Oct 2026, Day 1watchTowr shows what makes the read serious: Jira's WEB-INF/classes/crowd.properties holds application.name and, in plaintext, application.password for the Crowd directory. With them a caller reaches Crowd's REST API (its user search and user creation endpoints under /crowd/rest/usermanagement/1/) and can list users or create one. They note that a Crowd IP allow list makes this harder.source
15 Oct 2026, Day 10CISA adds the CVE to KEV with forensic triage required and a due date of 18 Oct 2026.source

The playbook's steps all fall in the first three days of that table. KEV came on day 10, so every step in the first 72 hours was taken seven days before CISA's listing and ten days before its due date. An organization that waited for the listing had three days to patch eight products on several release lines, and no early log search to show whether anyone had read a file in the ten days before. Neither Atlassian nor CISA gives an exploitation date, so that log search back to 5 Oct is the only way to date a read.

Sources

  1. Atlassian advisory CVE-2026-21589, read 19 Nov 2026
  2. watchTowr Labs on CVE-2026-21589, read 19 Nov 2026
  3. CISA KEV entry for CVE-2026-21589, read 19 Nov 2026
  4. NVD record for CVE-2026-21589, read 19 Nov 2026

Last reviewed 19 Nov 2026.