CVE-2026-90436: The 2006 Apache Axis Flaw Behind the Confluence and N-central Attacks

The CyberSec Guru

CVE-2026-90436

If you like this post, then please share it:

Buy me A Coffee!

Support The CyberSec Guru’s Mission

🔐 Fuel the cybersecurity crusade by buying me a coffee! Why your support matters: Zero paywalls: Keep the main content 100% free for learners worldwide.

“Your coffee keeps the servers running and the knowledge flowing in our fight against cybercrime.”☕ Support My Work

Buy Me a Coffee Button

Apache Axis 1.x shipped its last release in 2006 and is not getting another. The unauthenticated code-execution flaw disclosed in it on November 6, CVE-2026-90436, is now being exploited against Atlassian Confluence Data Center and N-able N-central. The HTTP 500 responses it leaves in your access logs look like failed attacks. Each one can be a completed exploit.

The first sign of trouble at a Rapid7 customer on November 13 was a burst of endpoint alerts for child processes a wiki server has no business spawning: Confluence’s Java process launching Bash and curl under the dedicated confluence service account. When Rapid7’s Managed Detection and Response team pulled the Confluence access logs, the matching requests looked like noise. They were unauthenticated POSTs to /plugins/servlet/soap-axis1/confluenceservice-v2, each answered with an HTTP 500 Internal Server Error. An analyst triaging by status code alone would have filed them under “broken client” or “scanner garbage.” Each 500 was an exploit that had already worked, and by the time the first one was logged the attacker had command execution on the host. Within twelve hours they had signed in to the victim’s Microsoft 365 tenant as an integration service account.

What is CVE-2026-90436?

CVE-2026-90436 is an unauthenticated remote code execution flaw in Apache Axis 1.x, a SOAP engine whose final release, 1.4, shipped on April 22, 2006. Axis parses a SOAP fault that arrives inside a request and instantiates whatever Java class the sender names. An attacker pairs that behavior with a gadget class already on the classpath and gets commands running as the service account.

The bug belongs to Apache Axis, not to Atlassian or N-able. The Apache Software Foundation labels the CVE “unsupported when assigned.” Confluence Data Center bundles Axis to power its legacy XML-RPC and SOAP remote APIs, and N-able N-central bundles it for its Legacy API. Both products keep the vulnerable surface behind a non-default administrative toggle, which is the only reason this is not a far larger emergency.

Apache credits José Apari-Pantigozo of OKIOK as the finder. The CVE was reserved on September 11 and published on November 6 under CWE-470 (unsafe reflection). Atlassian followed with its own advisory on November 9, tracking the work as CONFSERVER-104610.

How the Axis 1.x SOAP fault flaw works

SOAP’s fault mechanism carries errors from server to client. When a remote procedure call fails, the server returns a <Fault> element describing what went wrong, and Axis 1.x includes deserialization logic that rebuilds the matching exception object on the receiving side. According to Apache’s November 6 oss-security announcement, Axis applies that same reconstruction logic to faults arriving in inbound requests, with no check on direction. A client is not supposed to send a server a fault. Axis parses one anyway, and while doing so it loads and instantiates a Java class named by the sender.

Two classes do the work. As the SAX parser walks the <detail> block of an incoming fault, SOAPFaultDetailsBuilder reads the text of an <exceptionName> child and passes it to ClassUtils.forName(), an initializing class load. In Java, loading a class for the first time with initialization enabled runs its static {} block before any type check happens, so naming a class is enough to execute its static initializer.

If the detail block also carries a <data> element typed as xsd:string, SOAPFaultBuilder.createFault() goes further. It reflectively finds the class’s public single-String constructor and invokes it with the attacker’s data, and only afterwards tries to cast the result to AxisFault. In an attack the named class is never an AxisFault subtype, so the cast throws ClassCastException, a generic catch block swallows it, and Axis falls back to a plain fault. That fallback is the HTTP 500 the victim sees. The static initializer and the constructor have already finished running inside the JVM.

The trigger is an unremarkable SOAP envelope. Its body holds a Fault element, and the detail block inside it contains an exceptionName element naming a class and a data element carrying the string handed to that class’s constructor. The parser does not care whether the class resembles a fault.

Why the bug fires before authentication

Two properties make this a pre-authentication flaw. The fault is parsed while the Axis engine is still working out which service the request targets, before any per-service handler chain runs. Axis usually implements authentication as one of those handlers, so a service protected by an auth handler gets no protection from this bug. And because parsing happens during service resolution, the primitive is not limited to valid, deployed service operations.

📬 Stay Ahead of Cyber Threats

Get the latest cybersecurity news, critical vulnerabilities, threat intelligence, tutorials, and exclusive giveaways delivered straight to your inbox. No spam. Unsubscribe anytime.

From class instantiation to code execution: the gadget problem

Instantiating an arbitrary class with a string argument is a primitive, not a payload. To get code execution the attacker needs a gadget on the application’s classpath whose static initializer or String constructor does something useful to them.

On a bare Axis installation the primitive is still dangerous. Naming a JDK file-writing class with a filesystem path in <data> creates or truncates files owned by the service account, which is what the first public lab proof of concept, published November 10, demonstrates. Censys’s advisory notes that this lab PoC shows file creation and truncation, not a full command-execution demo against Confluence. I would keep that distinction in mind when judging exploit maturity.

The Confluence side: Spring and the ActiveMQ precedent

The escalation path is well trodden. Apache’s patch commit includes a regression test whose comments name Spring’s ClassPathXmlApplicationContext as the canonical dangerous gadget and compare the bug to CVE-2023-46604 in Apache ActiveMQ. Constructed with a URL string, that class fetches a remote Spring bean-definition XML and instantiates the beans it declares, which can spawn processes. The ActiveMQ exploitation wave used the same technique, and enterprise classpaths hold plenty of similar candidates.

The N-central gadget: JLine’s TerminalLineSettings

The N-central exploitation uses a subtler gadget, and it is not a deserialization classic. BleepingComputer’s reporting, corroborated by Previdian’s honeypot telemetry, identifies the class in the published chain as jline.internal.TerminalLineSettings from the JLine console library.

The JLine 2 source explains why it works. The public TerminalLineSettings(String ttyDevice) constructor treats its argument as a TTY device path and calls stty during initialization to read terminal settings. For any device string other than the default /dev/tty, the redirect fast path is disabled, and the code builds a command line of the form stty -g < <ttyDevice> and runs it through new ProcessBuilder(shCommand, "-c", …). The attacker-controlled device path is concatenated straight into a shell command, so shell metacharacters in that string become arbitrary command execution as the service account. No uploaded code and no second-stage deserialization are involved: a constructor shells out, and a parser calls constructors.

Rapid7 has not disclosed which gadget the Confluence attacker used. The observed Bash and curl child processes confirm that the Confluence classpath held something sufficient.

Apache’s patch, and why no fixed Axis 1.x release is coming

The patch commit changes three files. SOAPFaultDetailsBuilder now loads the named class without initialization through a new ClassUtils.forName(name, false) overload and accepts it only if AxisFault.class.isAssignableFrom(cls). SOAPFaultBuilder applies the same assignability check before invoking any constructor. A regression test asserts that a non-fault gadget class is neither initialized nor constructed.

The project says it does not expect to produce an Axis 1.x release with this fix. It recommends migrating to another SOAP engine such as Axis2/Java, or building Axis yourself with the patch applied. The Debian security tracker lists the distribution’s axis package as vulnerable with no fix.

The migration advice carries a footnote. In the same November 6 oss-security batch, Apache disclosed a separate set of Axis2/Java issues, including XXE via imported WSDL/XSD and SSRF via WS-Addressing endpoints. The successor is maintained, but it has its own attack surface.

Inside a Confluence intrusion: what Rapid7 saw in twelve hours

Rapid7’s published telemetry is the most detailed public account of this flaw in operation, and it reads like an automated credential-mining run more than a malware deployment. After the initial POSTs and their misleading 500s, the attacker ran hundreds of hex-encoded shell commands and pulled staging scripts from ephemeral tunnels on Pinggy and Cloudflare (*.trycloudflare[.]com), with callbacks to webhook[.]site. Each command wrapped its output in labeled delimiters and carried a sequential task ID. Rapid7 assesses that rigid framing as consistent with AI-assisted automated tooling that parses command output and dispatches follow-on tasks. That is Rapid7’s assessment, not a confirmed attribution, but the structure is distinctive enough to hunt for.

The pace was brisk. Just over three minutes after the first observed execution, the attacker was authenticating to the backend MySQL server with Confluence’s own database credentials, taken from the application’s configuration. Confluence keeps its JDBC settings, including the database password, in confluence.cfg.xml in the home directory. From there the attacker queried page bodies, application links, and attachments whose names suggested credentials, tokens, VPN configurations, and private keys. Scripts then swept Confluence’s default attachment store at /var/atlassian/application-data/confluence/attachments/v4/, parsed Office documents, and hunted with regular expressions for payment card numbers, using Luhn checksum validation to filter false positives before exfiltration.

Identity enumeration came next. The attacker used a Confluence application credential to query the internal Atlassian Crowd API for users and group memberships, then ran Impacket’s GetNPUsers.py against a domain controller in an attempted AS-REP roasting attack. Those attempts failed. The credential mining did not. About four hours after the first exploit request, the attacker held a service account credential and was querying a production cloud data warehouse. Roughly twelve hours in, they signed in to Microsoft 365 as an integration service account using legacy authentication from 192.238.253[.]176. That is the same address embedded in the hostname of the earlier Pinggy tunnel (wpefh-192-238-253-176.free.pinggy[.]net) and seen in the original exploit requests.

Almost everything the attacker gained after the foothold came from secrets stored inside Confluence, not from further exploitation. The unauthenticated RCE got them in, and the wiki’s contents were the prize. Every Confluence administrator should read that as a warning about what their own wiki holds.

Which Confluence Data Center versions are affected?

A checkbox decides exposure here, not a version string. Atlassian’s advisory lists every major branch from 1.x through 9.x as affected when the “Remote API (XML-RPC & SOAP)” setting is enabled. 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 unaffected. Atlassian Cloud sites (anything on atlassian.net or jira.com) do not run this architecture and are unaffected under any configuration.

The only fixed build the advisory names is 9.2.28, on the 9.2 Long-Term Support line that debuted in December 2024. Everything else resolves to “upgrade to a current release.” The exposed population is long-lived installations that upgraded incrementally from older versions, plus any where an administrator enabled the Remote API for legacy scripts, CI documentation publishers, or third-party add-ons.

CVSS scores and exposure numbers

The numbers need some care. Censys counted 1,212,568 internet-facing Confluence web properties around disclosure, excluding honeypots. A web property is a hostname-and-port pair, so the host-level count is lower, and a match shows only that Confluence is present. It says nothing about the version or the Remote API toggle.

The scoring has layers too. Apache published no numeric CVSS score, only a textual “critical.” The 9.8 that trackers repeat is CISA’s Authorized Data Publisher score under CVSS v3.1 (AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H). Atlassian’s internal assessment uses CVSS v4.0 and arrives at 9.3 (CVSS:4.0/AV:N/AC:L/AT:N/PR:N/UI:N/VC:H/VI:H/VA:H/SC:N/SI:N/SA:N). The gap reflects the two scoring methodologies and not a disagreement about the bug. NVD still lists the record as awaiting analysis.

Two enrichment records show how far automated triage can trail reality. CISA’s ADP SSVC assessment, timestamped November 10, recorded Exploitation: none, two days before miners were running on 21 N-central servers. As of November 17 the CVE had not been added to CISA’s Known Exploited Vulnerabilities catalog. Neither fact changes the urgency of a flaw Rapid7 has watched execute commands in a customer environment.

How N-central was exploited: 4EverAble2Shell and the XMRig miners

The N-central story is tangled up with a wider disclosure crisis. Tenable’s FAQ documents that a researcher operating as “Oracle-Security” on GitHub, since renamed “MSPNightmare” and identified by BleepingComputer as Sean Bales, published a write-up called “Able2Shell.” It describes 33 alleged N-central vulnerabilities with proof-of-concept code, and the earliest README commit is dated November 9 at 04:14 UTC.

N-able took its hosted N-central environments offline and urged on-premises customers to do the same. It then shipped N-central 2026.4 Hotfix 2 (build 2026.4.1.26) on November 10, fixing seven CVEs that span XXE, OGNL injection, SSRF, path-traversal RCE, and two local privilege-escalation paths. On November 12, Bales published two more exploit sets claimed to work against Hotfix 2: Agent2Shell, which targets Windows agents and probes for SYSTEM-level execution, and 4EverAble2Shell, an unauthenticated chain to root on the N-central server that requires the non-default Legacy API to be enabled.

4EverAble2Shell is where CVE-2026-90436 comes in. As BleepingComputer reported, the chain first uses the Axis flaw to run commands as the unprivileged nable service account. It then escalates through a separate N-central weakness. Previdian’s analysis describes abuse of the privileged setLDFileOwner helper and a symbolic link to make a normally root-owned script writable, followed by editing that script and invoking N-central’s serverControl functionality to execute as root, with an attempt to restore the files afterward. Previdian tracks that escalation as VULN-2026-0009. The Axis step alone does not give root. The full chain does.

Exploitation timeline and the XMRig payload

Exploitation began within hours of the PoC. Previdian saw probing against its N-central honeypot at roughly 08:30 UTC on November 12, using the same Java class referenced in the published exploit, followed by full-chain attempts. Huntback saw the same source IP probe at 08:30 and send the complete chain at 09:16. The payload downloaded a shell script from 157.20.241[.]16 that installed an XMRig Monero miner configured to use about half the available CPU cores, presumably to blunt detection.

Huntback then enumerated victims through the attacker’s mining-pool statistics, where workers were named nc_ plus the server IP. At 12:50 UTC it counted 21 active miners, all on hosts serving N-central’s web interface, and 19 of them reported build 2026.4.1.26, the hotfix shipped two days earlier. The compromised servers sat in the United States, Canada, South Africa, the United Kingdom, Germany, France, and Sweden. Host artifacts include an unexpected xvd.service systemd unit and files under /tmp/.xvd or /var/tmp/.xvd.

N-able published network IOCs on November 13 for customers who had not disabled the Legacy API. Rapid7 lists 157.20.241[.]16, 139.99.69[.]109, and 79.76.51[.]56 among them.

N-central 2026.4 HF3: the vendor response

The vendor response culminated on November 16 with N-central 2026.4 HF3 (build 2026.4.1.39). It addresses seventeen new CVEs (CVE-2026-129271 through CVE-2026-129287) and removes the Legacy API configuration option entirely, so the feature is now permanently blocked instead of merely off by default. The release also hardens agent and probe authentication. N-able says all customers, hosted and on-premises, must upgrade agents and probes to 2026.4.1 HF3 or remain open to agent impersonation. Per N-able’s remediation update, the Intune integration and built-in remote control (other than Take Control) stay disabled pending further remediation.

For managed service providers the stakes are structural. An RMM appliance holds the management plane for every downstream customer, so root on that server is a potential multi-tenant incident and not a single-server one.

How to fix or mitigate CVE-2026-90436

Confluence Data Center

The durable fix is to upgrade to 9.2.28 or a current unaffected 10.x release. Organizations on the 1.x through 8.x branches have no named in-place fix and need to plan a migration.

Where patching has to wait, Atlassian’s interim mitigation is to disable the legacy remote API outright:

  1. Open General Configuration, then Further Configuration. You can reach it directly at <base_url>/admin/viewspacesconfig.action.
  2. Click Edit.
  3. Clear the “Remote API (XML-RPC & SOAP)” checkbox.
  4. Save.
  5. Verify with a single unauthenticated request: browse to or curl -i the endpoint at <base_url>/plugins/servlet/soap-axis1/confluenceservice-v2.

A correctly applied mitigation returns HTTP Status 403 – Forbidden. If you instead get a message that references confluenceservice-v2 or an “AXIS service,” the servlet is still reachable and the setting has not taken effect. The change applies immediately without a restart, but scripts, automations, and add-ons that depend on the legacy API may stop working, so Atlassian recommends testing in non-production first.

Compensating controls at the proxy or WAF

The flaw triggers during XML deserialization, before Axis invokes service-level authentication handlers. Confluence authentication plugins and session-based reverse-proxy checks that allow unauthenticated access to the servlet path therefore will not stop it.

Where a reverse proxy or WAF sits in front of Confluence, block external requests to /plugins/servlet/soap-axis1/* and to the XML-RPC endpoint at /rpc/xmlrpc. If body inspection is available, also drop inbound SOAP bodies that carry a <Fault> with an <exceptionName> child. These are stopgaps and do not replace removing the endpoint.

N-central

Upgrade to 2026.4 HF3 and upgrade every agent and probe. HF3 removes the Legacy API toggle, so verification is that the option is gone and that /dms/services/ServerUI no longer answers. Requests to that path that reference jline.internal.TerminalLineSettings are a direct exploitation tell.

How to detect CVE-2026-90436 exploitation

Successful exploitation masquerades as a server error, so my advice is to stop triaging these requests by status code. Start in Confluence’s Tomcat access logs, the conf_access_log.* files under the installation’s logs directory. Look for unauthenticated POSTs to /plugins/servlet/soap-axis1/* returning HTTP 500 (or 200). Then correlate those timestamps, within a tight window, against EDR process-creation events where the parent is the Confluence java process and the child is bash, sh, curl, wget, or python. In the Rapid7 case the child processes appeared within seconds of the requests.

From there, widen the search:

  • Database audit logs for interactive MySQL authentication from the Confluence host outside normal connection-pool behavior, and for bulk selects against page-body and attachment tables.
  • File-access telemetry for sequential reads across the attachments/v4/ tree.
  • Crowd access logs for bulk user and group enumeration through the Confluence application account.
  • Active Directory security event 4768 with pre-authentication type 0, which points to AS-REP roasting attempts from the Confluence server.
  • Microsoft Entra sign-in logs for service-account authentications over legacy protocols.

Indicators of compromise

Rapid7’s Confluence-side network indicators are 192.238.253[.]176, 5.133.47[.]243, 163.182.68[.]109, and 104.13.106[.]166, along with wpefh-192-238-253-176.free.pinggy[.]net, *.trycloudflare[.]com, and webhook[.]site. A hit warrants immediate investigation. A miss proves nothing, because the attacker relied on ephemeral tunnels, the infrastructure rotates, and IP-based detection alone falls short.

For N-central, pair the network indicators above (157.20.241[.]16, 139.99.69[.]109, 79.76.51[.]56) with the host artifacts: the xvd.service unit, the .xvd directories, and unexpected mining processes. Treat any confirmed hit as a full root compromise of the appliance.

Atlassian says it cannot confirm whether a given instance was affected and directs customers to their own security teams. The same applies to any exposed Axis endpoint.

Why a 20-year-old library is still a supply-chain risk

Strip away the product names and this is a story about dependency retention. Axis 1.x was superseded by Axis2 two decades ago, yet axis.jar still ships inside commercial Java archives to serve optional “Legacy API” checkboxes. Vulnerability scanners that evaluate only top-level product versions rarely inspect nested JARs and runtime servlet mappings. Turning on one convenience toggle can put a twenty-year-old XML parser in front of the internet, and the only durable remedy is to remove or replace the dependency. That is what Atlassian did in Confluence 10.0 and what N-able did in HF3.

The immediate priorities are concrete. Patch Confluence to 9.2.28 or later, or disable the Remote API where you cannot. Move N-central to 2026.4 HF3 and upgrade every agent and probe. Hunt for indicators of exploitation that has already happened. Then run software-composition analysis and filesystem sweeps across your Java estate for org/apache/axis/message/SOAPFaultBuilder.class, and treat any Axis 1.x binary without commit 41f3b4e8 as an unauthenticated RCE liability waiting for a gadget.

CVE-2026-90436 FAQ

Is CVE-2026-90436 a Confluence vulnerability?

No. It is a flaw in Apache Axis 1.x. Confluence Data Center is exposed because it bundles Axis for its legacy XML-RPC and SOAP remote APIs, and only when the “Remote API (XML-RPC & SOAP)” setting is enabled.

Is Atlassian Cloud affected?

No. Atlassian Cloud sites on atlassian.net or jira.com do not run this architecture and are unaffected under any configuration.

Which Confluence version fixes CVE-2026-90436?

The only fixed build Atlassian’s advisory names is 9.2.28 on the 9.2 Long-Term Support line. Confluence 10.0 removed the Remote API feature, so 10.x is unaffected. Other branches need an upgrade to a current release.

Does an HTTP 500 mean the exploit attempt failed?

No. The 500 comes from Axis’s fallback after the class cast fails, and the static initializer and constructor have already run by then. Rapid7 saw HTTP 500 responses on requests that gave the attacker command execution.

Is there a patch for Apache Axis 1.x itself?

Apache has committed a fix, but the project does not expect to produce an Axis 1.x release containing it. It recommends migrating to another SOAP engine such as Axis2/Java or building Axis yourself with the patch applied.

Does the N-central exploit give root access?

The Axis flaw alone runs commands as the unprivileged nable service account. The 4EverAble2Shell chain then escalates to root through a separate N-central weakness that Previdian tracks as VULN-2026-0009. N-central 2026.4 HF3 removes the Legacy API option entirely.

Buy me A Coffee!

Support The CyberSec Guru’s Mission

🔐 Fuel the cybersecurity crusade by buying me a coffee! Your contribution powers free tutorials, hands-on labs, and security resources.

Why your support matters:
  • Writeup Access: Get complete writeup access within 12 hours
  • Zero paywalls: Keep the main content 100% free for learners worldwide

Perks for one-time supporters:
☕️ $5: Shoutout in Buy Me a Coffee
🛡️ $8: Fast-track Access to Live Webinars
💻 $10: Vote on future tutorial topics + exclusive AMA access

“Your coffee keeps the servers running and the knowledge flowing in our fight against cybercrime.”☕ Support My Work

Buy Me a Coffee Button

If you like this post, then please share it:

Exploits

Discover more from The CyberSec Guru

Subscribe to get the latest posts sent to your email!

Leave a Comment

This site uses Akismet to reduce spam. Learn how your comment data is processed.