9.3

CVSS 4.0 score Atlassian gives the Confluence flaw

3 min

from first code execution to the Confluence database, per Rapid7

12 hours

from the first request to a Microsoft 365 sign in

21 Nov

federal remediation deadline set by CISA

A library with no fix makes every vendor the patch

CVE-2026-90436 is a bug in a library that nobody maintains, and that changes who has to fix it. Apache has said plainly that it does not expect to ship a fixed Axis 1.x release, so the only patches that will ever exist are the ones each vendor builds into its own product.

That is the part of this week's news that will outlast the Confluence emergency. Apache Axis 1 is a Java SOAP engine from the early web services era. It stopped receiving releases long ago, but it never stopped being shipped. It sits inside enterprise products that grew SOAP interfaces in the 2000s and kept them for compatibility, and a lot of those products do not list it anywhere a customer would look. When the library itself has a critical flaw and no maintainer, the advisory you need is not Apache's. It is the one from every vendor that bundled the jar, and some of those vendors will be slow, some will be gone, and some will not know they ship it.

Confluence Data Center is the first product where the exploitation is public, and it shows how much is at stake. In the intrusion Rapid7 described, a wiki server became a path to the company's database, its identity directory, a production cloud data warehouse and Microsoft 365 within a working day. Nothing in that chain needed a second bug. It needed the credentials a wiki collects over ten years.

So the useful question this week is two questions. Is any Confluence server of ours still running the Remote API? And what else in the estate carries Axis 1 inside it? The first has a clean answer by tonight. The second is an inventory problem, and it is the real work.

What Apache, Atlassian and CISA say about CVE-2026-90436

The Apache Axis announcement of 6 Nov 2026 rates the flaw critical and marks it unsupported when assigned. It affects all Apache Axis 1.x versions through 1.4. In Apache's description, an inbound SOAP fault can name a class, and Axis loads that class before checking whether it is the kind of fault it expected. Loading a class runs its static initializer, and with a suitable class on the application's classpath that leads to remote code execution, with no authentication. Apache recommends moving to a different SOAP engine such as Axis 2, or running a build of Axis 1 with a single upstream commit applied.

Atlassian published its advisory on 9 Nov 2026. Confluence Data Center from 1.x through 9.x is affected when the setting called Remote API (XML-RPC and SOAP) is switched on. That setting has been off by default for new installations since Confluence 3.0, and the feature was removed entirely in Confluence 10.0, so 10.x is not affected. Atlassian rates it critical, 9.3 under CVSS 4.0, and lists 9.2.28 as the fixed release. Atlassian Cloud sites are not affected. For anyone who cannot upgrade, the workaround is to clear the Remote API checkbox under General Configuration, Further Configuration, with a warning that scripts, automations and third party add-ons that depend on it may stop working. Atlassian also says it cannot confirm whether a given instance has been affected.

On 18 Nov 2026 CISA added CVE-2026-90436 to the Known Exploited Vulnerabilities catalog as the Apache Axis Unsafe Reflection Vulnerability, based on evidence of active exploitation. The catalog entry lists the weakness as CWE-470, gives federal civilian agencies until 21 Nov 2026, flags the entry for forensic triage, and carries a note that matters more than the deadline: the flaw could affect an open source component or third party library used by different products. CISA is naming the library, not the wiki.

CVE-2026-90436 Critical
  • Component //

    Apache Axis 1.x through 1.4, end of life
  • Confluence affected //

    Data Center 1.x to 9.x with the Remote API (XML-RPC and SOAP) on
  • Fixed in //

    Confluence 9.2.28, or 10.x where the feature is gone; no fixed Axis 1.x release expected
  • CVSS //

    9.3 (4.0, Atlassian)
  • Needs //

    network reach to the SOAP endpoint, no credentials
  • Exploited //

    Yes, per CISA KEV, added 18 Nov 2026

The exploitation record is unusually detailed. Rapid7 reports that its managed detection team saw the flaw exploited against Confluence Data Center in a customer environment on 13 Nov 2026, a week after Apache's disclosure. The same write-up notes that a public proof of concept aimed at N-able N-central appeared on 12 Nov, and that a public Metasploit module for Confluence followed on 17 Nov. Once a framework module exists, the population of people who can run this is no longer small.

Sensor data points the same way. Previdian counted 51 exploitation attempts from 14 addresses between 12 Nov and 19 Nov 2026, and is careful to say its sensors do not show whether any target was vulnerable or whether an attempt succeeded. That is background scanning, not a victim count, and we read it that way. SecurityWeek's report carries the same timeline and the federal deadline independently.

How one SOAP request reached Microsoft 365 in 12 hours

The Rapid7 account is worth reading as a map of what a wiki server is connected to. We have kept it at the level of what each stage achieved.

Figure 01

From the wiki to the tenant in one day

  1. 01

    In //

    One unauthenticated SOAP request runs code as Confluence.

  2. 02

    Read //

    Config, keys and the database password.

  3. 03

    Mine //

    Pages and attachments searched for secrets.

  4. 04

    Reach //

    Directory, data warehouse, service accounts.

  5. 05

    Out //

    A Microsoft 365 sign in as an integration account.

Result //

A wiki became the identity of the business.

No second vulnerability was needed. Every step after the first used credentials the wiki already held.

The entry is a POST to the Confluence SOAP endpoint. Rapid7 notes that the server answered with HTTP 500 errors while the code still ran, which matters for anyone reading logs: a failed looking request is not a failed attack. Endpoint telemetry then showed the Confluence Java process launching a shell and a download tool as the confluence service account within seconds.

What followed was fast and methodical. Rapid7 describes hundreds of hex encoded shell commands delivered through short lived tunnelling domains, in a structure it says suggests AI assisted automated tooling. The attacker enumerated the host, the network, the Confluence configuration, its encryption keys and its database settings. About 3 minutes after the first command ran, the attacker logged in to the MySQL database with Confluence's own credentials and queried page bodies, attachments and application links.

Then came the part every Confluence owner should sit with. The attacker searched attachments for credentials, VPN configurations, private keys and payment card numbers. That is a fair description of what accumulates in a company wiki: runbooks with passwords pasted in, exported VPN profiles attached to onboarding pages, keys dropped in for a colleague one afternoon in 2019. The attacker also used a Confluence application credential to query the Atlassian Crowd API, which is the user directory many Atlassian estates hang off, and tried to request crackable Kerberos material from a domain controller for accounts without preauthentication. Rapid7 says that last attempt failed.

It did not need to succeed. About four hours after the first exploit request, the attacker had a service account credential and was querying a production cloud data warehouse. About 12 hours after it, the attacker signed in to Microsoft 365 as an integration service account using legacy authentication. Rapid7 does not attribute the activity, and neither do we.

The order of that chain is the lesson. The attacker went to the database before the network, to attachments before Active Directory, and to service accounts before people. Wikis, ticketing systems and build servers hold the credentials that integrations run on, and integration accounts are the ones nobody enrolled in phishing resistant MFA because nobody logs in as them by hand. We made the same argument about management planes in our N-central write-up, which is also on the Axis 1 list: whatever a system is trusted to reach, an attacker on that system reaches too.

Why authentication on the service does not help

One sentence in Apache's announcement explains why this flaw is worse than an ordinary SOAP bug. The fault is parsed while the engine works out which service the request is for, before any per-service handler chain runs, so authentication handlers configured on a service do not protect against it.

Most teams that still expose an old SOAP interface have made peace with it on the grounds that it sits behind a login. On a typical web services stack, authentication is a handler on the service: a request arrives, the handler checks the credentials, and only then does the business code run. CVE-2026-90436 fires earlier than that. The attacker does not need to pass the handler because the vulnerable code runs on the way to it. Any argument of the form "the endpoint is authenticated, so it is low risk" does not apply to this bug.

The second reason is the absence of an upstream fix. With a maintained library, the advisory cycle is familiar: upstream releases a version, scanners learn the version string, and dependency tools flag every product that ships the old one. None of that happens here. Apache has published a commit, not a release. Products that rebuild Axis 1 with the commit applied may still report version 1.4 in the jar's metadata, and products that have not rebuilt it will report exactly the same thing. A version based scanner cannot tell a patched Axis 1 from a vulnerable one, so the vendor's own advisory and fixed release are the only reliable signal.

That is why the inventory matters more than the patch. Confluence has a fix and a clear setting. N-able's answer, in its 2026.4 HF3 release notes of 16 Nov 2026, was to remove the configuration option and block the legacy APIs permanently. Other vendors that bundle Axis 1 have not necessarily said anything yet, and an Axis 1 jar in a product's lib directory does not announce itself. Treat CISA's note about third party libraries as a prompt to go and look, not as background.

How to tell if your Confluence server was hit

Start with exposure, because it decides how much of the rest you need. If the Remote API setting was off on every Confluence Data Center server for the whole period since 6 Nov 2026, this flaw was not reachable through Confluence. If it was on, assume the server was probed, and look for evidence that a probe worked.

First, the web logs. Rapid7 names the endpoint as the soap-axis1 servlet path under /plugins/servlet/, and notes that unauthenticated POST requests to it returned HTTP 500 during exploitation. Search the access logs and any proxy or load balancer logs in front of Confluence for POSTs to that path from addresses that are not your known integrations. A burst of 500 responses there is a lead, not noise.

Second, the process tree. Confluence is a long running Java process under its own service account. A shell, a download tool or a scripting interpreter launched from it is rare on a healthy server and is the clearest trace this intrusion left.

Figure 02

Confluence spawns a shell or download tool

Sigma

title: Unexpected child process of Confluence
status: experimental
description: A shell, interpreter or download tool started by the Confluence JVM, the trace Rapid7 observed after exploitation of CVE-2026-90436
logsource:
  product: linux
  category: process_creation
detection:
  parent:
    ParentImage|endswith: '/java'
    ParentCommandLine|contains: 'confluence'
  child:
    Image|endswith:
      - '/sh'
      - '/bash'
      - '/dash'
      - '/curl'
      - '/wget'
      - '/python3'
      - '/perl'
      - '/base64'
      - '/xxd'
  condition: parent and child
falsepositives:
  - Add-ons or maintenance scripts that shell out by design; baseline them and filter by command line
level: high
A starting point to tune against your own telemetry, not a production rule. On Windows hosts, swap the child list for cmd.exe, powershell.exe and similar.

Third, the database. The attacker in Rapid7's case logged in to MySQL with Confluence's own credentials. If database audit logging is on, look for sessions from the Confluence account that ran broad reads of page content, attachment tables or application links outside the application's usual query pattern, and for any session from that account originating somewhere other than the application host.

Fourth, identity, which is where the damage lands. Pull every credential the Confluence server could read: the database password, application link secrets, Crowd application credentials, mail and LDAP binds, and any service account whose password or key appears in a page or attachment. Then check their recent use. Sign ins by integration accounts from new addresses, legacy authentication where none is expected, and queries to data platforms from accounts that normally only write to them are the signals that matter. Our Learn guide on reviewing Entra sign in logs covers how we read those records for service accounts and legacy protocols.

If any of the first three checks hits, treat the server as compromised and rotate everything in the fourth list, not just the database password. CISA's forensic triage flag on this entry points the same way. Rebuilding the host without rotating the secrets it held leaves the attacker holding the keys to the systems the wiki talked to.

How we find retired libraries on an engagement

On internal and external engagements we look for old web services surfaces early, because they tend to be the least watched code on the network. SOAP and XML-RPC endpoints often survive inside products whose main interface was modernised years ago, and their presence is usually visible from the outside: a servlet path, a WSDL document, a service listing page, a characteristic fault response.

At a high level, on a sanctioned engagement, the method has three parts. We map every application that answers on a legacy web services path, internal and internet facing, and match each one to the vendor and product behind it. We then read what the product actually ships, from its documentation, its install tree where we have access, or the vendor's own component lists, to find bundled libraries that no longer receive releases. Finally we test whether the endpoint processes untrusted input before authentication, which is often the real finding regardless of which CVE is current. Our Learn guide on exposed admin and debug endpoints describes the read without using approach we take to these surfaces.

Where the scope allows it, we show impact without damage. On a system like Confluence that usually means demonstrating what the service account can reach, such as a credential found in an attachment or an application link with broad rights, rather than running anything against production data. The report then rates the finding by what the server holds and connects to, not by the age of the library. An end of life jar is a hygiene note on its own. An end of life jar parsing untrusted input on a server that holds the cloud warehouse credentials is the critical finding, and the report should say so in those words.

What is still coming

Three things are worth watching over the next few weeks.

More products on the list. CISA has already flagged the entry as a library issue, and Confluence and N-central are only the two products with public exploitation. Expect further vendor advisories as product teams check their own builds, and expect some products to stay silent because the vendor has stopped issuing updates. Ask your vendors directly whether they ship Apache Axis 1, and record the answer.

Wider tooling. A Metasploit module and a public proof of concept already exist, and the attacker in Rapid7's case used automation from the first request. Exposure windows for this class of bug are now measured in days. Any Confluence server that still answers on the SOAP path after this week should be treated as compromised until proven otherwise, not as merely unpatched.

The same shape elsewhere. The same week, SecurityWeek reported that Apache says Roller, its Java blogging server, is being actively exploited through CVE-2026-129448, a deserialization flaw in its XML-RPC interface that is processed before authentication and is fixed in Roller 6.1.6. Different product, same lesson: protocol endpoints built for machine clients in another decade are still parsing untrusted input on the network, and they are where attackers are looking this month. We made a related point about how fleet tools decide who is trusted in our Puppet analysis this week.

What to do today

  1. Close the Confluence path tonight. Upgrade Data Center to 9.2.28 or to 10.x, or clear the Remote API (XML-RPC and SOAP) setting until you can. Confirm on every node, including test and disaster recovery instances.

  2. Block the SOAP endpoint at the edge. No Confluence or N-central legacy API path should answer from the internet, whatever the patch level.

  3. Hunt on any server where the setting was on. Search access logs for POSTs to the soap-axis1 path, run the Sigma rule above, and review database sessions from the Confluence account.

  4. Rotate what the wiki held. Database and application link credentials, Crowd and LDAP binds, and every secret found in pages or attachments, then review recent sign ins by those service accounts.

  5. Find the rest of Axis 1. Search install trees and software bills of materials for axis 1.x jars, and ask each vendor in writing whether its product bundles Axis 1 and when it will ship a fix.

  6. Move secrets out of the wiki. Search pages and attachments for keys and passwords, replace them with links to a vault, and treat the wiki as holding whatever it held until you have done so.