Attackers are using an Apache Axis 1 flaw, and Apache doesn't plan a fixed release
A crafted SOAP error message makes end-of-life Apache Axis 1 run code before per-service login checks run. Here's who is affected, what Atlassian and N-able fixed, how to find hidden axis.jar copies, and which credentials to rotate.
part 46 of 47 · Platform releases, explained from scratchthe ideaAxis reads the fault before per-service login checks run.
after reading, you can- Find every copy of Axis 1 you run, even the ones a vendor tucked into a WAR file
- Fix Confluence and N-central, then check that the fix actually took
- Work out which stored passwords an attacker would go for next, and change them first
A SOAPdefinitionAn older way for systems to exchange XML messages over HTTP. Lots of enterprise integrations still use it, even where newer apps have moved to REST. library that stopped getting updates years ago now has a flaw attackers are using, and CISA added it to its exploited list today. If anything you run still carries Apache Axis 1.x, there's no fixed release coming from Apache. You patch through the product vendor, switch the SOAP endpoint off, or move off Axis. Confluence Data Center and N-able N-central are the two known targets so far.
An old piece of plumbing inside many Java apps will run code for anyone who knows how to ask.
Lots of business software talks to other systems by passing XML messages back and forth, a style called SOAP. Apache Axis 1 was a popular free toolkit for that, and vendors built it into their products. Its last release was 1.4, and the project calls it end of life. The new bug sits in the part that reads error messages. Send it a crafted error and it loads and starts a program component the sender names. Nobody has to log in first. With the right components already sitting in the application, the sender ends up running their own commands on your server.
It's like a front desk that takes complaint forms from anyone off the street. One box says which staff member should deal with it, and the clerk calls that person out before reading anything else. Write down whoever holds the master keys and they walk into the lobby carrying them. After the fix, the clerk only calls people from the complaints team.
- Confluence Data Center with the remote API on
- N-central servers with the Legacy API on
- Older Java apps that offer SOAP services
- Vendors whose products bundle axis.jar
- Integration accounts that sign in to M365
- US federal agencies and their contractors
Why it matters to you
Most patch news gives you a version to install. This one doesn't. Apache says it won't ship an Axis 1.x release with the fix, so every product carrying the library needs its own answer, and some will never get one.
CISA listed the flaw on its Known Exploited Vulnerabilities catalog, the list of bugs it has evidence of attacks against. The entry notes the bug could affect an open-source component used by many different products. It also flags the affected product as possibly end of life and tells users to stop using it or move to something supported. Federal civilian agencies have three days, and CISA also marks the entry for forensic triage, so they're expected to look for past compromise as well.
The attack Rapid7 documented shows why a SOAP bug turns into an identity problem fast. Seconds after the first request, Confluence's own Java process started running shell commands. Within minutes the attacker used Confluence's database password. By the twelve-hour mark they had signed in to Microsoft 365 as an integration service account, using legacy authenticationdefinitionOld sign-in protocols that take a username and password and skip multifactor. Microsoft says to block them with Conditional Access.. The code flaw was the way in. The stored credentials were what let them spread.
-
01 A crafted SOAP fault
The attacker posts a SOAP message to an Axis endpoint. It looks like an error report, but there's a class name hidden in its detail section.
-
02 Parsed before the login check
Axis reads the message while it's still working out which service it's for, so the service's own authentication hasn't run yet.
-
03 Axis loads the named class
The old code loads whatever class was named, which already runs that class's start-up code, and then builds an object from it.
-
04 A gadget does the rest
If a suitable class already sits on the classpath, building it runs the attacker's commands as the app's own account.
Under the hood, CISA files it as CWE-470, unsafe reflection. Axis looks for an exceptionName element in the fault's detail section and loads whatever class it names. If there's also a String data element, Axis calls that class's String constructor, and only afterwards checks whether it was an AxisFault. The fix commit names Spring's ClassPathXmlApplicationContext as a working gadgetdefinitionCode that's already in the app and does something risky the moment it's created. Attackers use a flaw like this one to get it created on their terms., the one used against ActiveMQ in CVE-2023-46604. So if an app has Spring on the classpath next to Axis 1, treat it as exploitable today.
| Who | Rating | What they say |
|---|---|---|
| NVD (CVSS 3.1 from CISA) | 9.8 Critical | Network, no login, no user action |
| Atlassian | 9.3 Critical (CVSS 4.0) | Patch Confluence Data Center to 9.2.28 or move to 10.x |
| Apache Axis project | Critical | Migrate to another SOAP engine such as Axis2/Java, or build Axis with the fix commit |
| CISA KEV | Exploited | Apply vendor mitigations or stop using the product. Due November 21 |
| Debian | Unfixed | Marked as a minor issue and ignored for bookworm and trixie |
| Ubuntu | Needs evaluation | Every supported release's axis package still awaiting triage |
Don't read the Linux distribution entries as reassurance. Debian's tracker lists its axis package as vulnerable in every release, with the stable ones set aside as a minor issue, and Ubuntu hasn't triaged its copies yet. Those labels are about the distribution package. A vendor's product bundling its own axis.jar won't show up in either tracker, and that's where the attacks are.
What got fixed, and what never will be
There are three kinds of fix here, and only the first two come from someone else. Atlassian and N-able have shipped product fixes. Apache published a code change but no release. Everything else is on you: finding copies of Axis 1 and either switching them off or replacing them.
- Loads whatever class the fault names, no matter who sent the message
- Just loading the class runs its start-up code. Nothing has been checked yet
- Builds the object first and asks what it is afterwards, when the damage is done
- The error is swallowed inside Axis, though the access logs still show an HTTP 500
- Still loads the named class, but its start-up code doesn't get to run
- Takes the class only if it's really an AxisFault, and drops anything else
- Won't build an object from a class that isn't a fault. A gadget never gets started
- Real custom faults still work, and a regression test covers the bug
Where a fix exists today
Atlassian's fixed release for the 9.x line. Confluence 10.0 removed the remote API entirely, so 10.x isn't affected. Atlassian Cloud sites aren't affected.
N-able's critical hotfix takes away the legacy API setting, so nobody can turn those APIs back on. N-able upgrades hosted servers itself.
The code fix in the Axis 1 repository. Apache suggests it only for teams that build their own Axis. There's no 1.x release that includes it.
End of life and marked unsupported when the CVE was assigned. Apache's advice is to move to another SOAP engine such as Axis2/Java.
Turning off Remote API (XML-RPC and SOAP) closes the hole without a restart, though scripts and add-ons that use it will stop working.
Still vulnerable or awaiting triage. Neither Debian nor Ubuntu has shipped a patched package so far.
| What you run | Fix or safe state | Watch out for |
|---|---|---|
| Confluence Data Center 1.x to 9.x | 9.2.28, or any 10.x | Only exposed with Remote API (XML-RPC and SOAP) on. It's been off by default for new installs since 3.0 |
| N-able N-central (on-premises) | 2026.4 HF3, build 2026.4.1.39 | Upgrade every Agent and Probe too, or they can be impersonated |
| N-able N-central (hosted) | Upgraded by N-able | You still have to upgrade Agents and Probes |
| Your own app on axis:axis 1.4 | Migrate to Axis2/Java or JAX-WS | A patched build of 1.4 from the fix commit is a stopgap, not a plan |
| A vendor product that bundles axis.jar | Ask the vendor | No statement yet means unknown, not safe |
From mailing list to KEV in twelve days
- Nov 6, 2026Apache publishes the CVE
The advisory goes to the Axis user list, marked unsupported when assigned, with the fix commit and no release.
- Nov 9, 2026Atlassian advisory and 9.2.28
It covers Confluence Data Center 1.x to 9.x, but only where the remote API is enabled.
- Nov 12, 2026Public exploit code for N-central
Rapid7 says a proof of concept targeting N-central appeared as part of a run of public zero-day disclosures.
- Nov 13, 2026Attacks on both products
Rapid7 sees Confluence exploited at a customer. N-able reports attacks on N-central, which Rapid7 ties to this CVE.
- Nov 16, 2026N-central 2026.4 HF3
The legacy APIs are blocked for good.
- Nov 18, 2026CISA adds it to KEV
Federal due date set three days out.
- Nov 21, 2026Federal due date
Twelve days from disclosure to the KEV list is quick for a bug in a library this old. The awkward part is everything else that bundles Axis 1. Those products may be affected too, and so far we haven't found a statement from any of their vendors.
Confluence admins can check the mitigation themselves. Atlassian says that once the remote API is off, browsing to the soap-axis1/confluenceservice-v2 path should return HTTP 403 Forbidden. Any page that still mentions confluenceservice-v2 or an AXIS service means the setting didn't take. It takes a minute, and it's the fastest proof you'll get before an upgrade window.
N-able went further than patching. Its hotfix removes the legacy API option altogether, and N-able temporarily turned off the Intune integration while it added more protections. Hosted customers got protections applied for them. Everyone, hosted or on-premises, still has to upgrade the Agents and Probes, which N-able says can otherwise be impersonated.
What to do now
What to do now, by who you are
- Ask IT whether you run Confluence Data Center or N-central yourselves, and when each was last upgraded.
- Ask every vendor whose Java software you host where they stand on CVE-2026-90436. Get the answer in writing.
- If you sell to US federal agencies, expect a question about the November 21 date. Have your answer ready.
- Ask who is checking for past compromise. Patching today doesn't tell you about last week.
- Upgrade Confluence Data Center to 9.2.28 or a current 10.x release. If you can't today, turn off the remote API and check for the 403.
- Move on-premises N-central to 2026.4 HF3, then upgrade every Agent and Probe.
- Block the SOAP path at the reverse proxy for any app where nobody can say who uses it.
- Keep the admin and API ports of these servers off the public internet.
- Search builds for axis:axis and org.apache.axis:axis. Look in fat jars and WAR files as well, since shaded copies hide there.
- Plan a move to Axis2/Java or JAX-WS for services you own. Apache has a migration guide from Axis 1.
- If you must keep Axis 1 for a while, build it with commit 41f3b4e and pin that build.
- Check whether Spring or another known gadget library sits on the same classpath. Those apps go to the front of the queue.
- Hunt for unauthenticated POSTs to Axis endpoints that returned HTTP 500, then for the app's Java process starting a shell.
- Write down every credential stored on the exposed servers. That's database passwords, Crowd and directory bind accounts, and integration logins.
- Check Entra sign-in logs for legacy authentication from service and integration accounts.
- Look for Kerberos requests aimed at accounts that don't require pre-authentication.
If you only have an hour, check internet-facing systems first. A Confluence or N-central server reachable from outside with the remote or legacy API on is the case attackers already have a script for. Internal servers come next. Rapid7 recommends checking those too, since anyone already inside your network can reach them.
# example output with made-up paths. Run on app servers and inside images, since each one can differ find / -xdev -iname 'axis*.jar' -not -iname 'axis2*' 2>/dev/null /opt/vendorapp/WEB-INF/lib/axis-1.4.jar find / -xdev \( -name '*.jar' -o -name '*.war' \) -print0 2>/dev/null | while IFS= read -r -d '' j; do unzip -l "$j" 2>/dev/null | grep -qE 'org/apache/axis/client/Call\.class|WEB-INF/lib/axis(-1[^/]*)?\.jar' && echo "$j"; done /opt/billing/billing-service.war # Axis 1 was bundled inside this WAR, so a search by file name alone would have missed it mvn dependency:tree -Dincludes=axis:axis,org.apache.axis:axis [INFO] \- axis:axis:jar:1.4:compile curl -so /dev/null -w '%{http_code}' $WIKI/plugins/servlet/soap-axis1/confluenceservice-v2 # you want 403 403 # 403 is what Atlassian says you'll get once the remote API is really off
An order that gets Axis 1 out of the line of fire
- 01Fix the two known targets
Confluence Data Center and N-central come first, because public attacks already exist for both. Patch them, or switch the API off and confirm it's off.
- 02Build the list
Search servers, images and build files for Axis 1 jars. Look inside WAR and fat jar files too, because shaded copies don't show up by name. SBOMs help, but check what's actually running.
- 03Rank by who can reach it
Anything that accepts SOAP from the internet or from partners goes to the top. Anything with Spring or another known gadget library on the same classpath comes next.
- 04Close the door while you wait
Block the Axis servlet path at the proxy for apps nobody uses SOAP on. Where a partner needs it, limit it to their addresses.
- 05Get vendor answers
Write to each vendor whose product bundles axis.jar and ask for a fix date or a mitigation. Record the reply, or the silence.
- 06Replace it
For code you own, move the services to Axis2/Java or JAX-WS. A patched 1.4 build buys time, but it's still an unmaintained library.
Harden so the next one hurts less
This bug is a reminder about two kinds of debt. One is old libraries nobody tracks. The other is what an app's service account can reach once someone controls the app. The attacker Rapid7 tracked spent far more time on the second than the first.
Before: an old SOAP endpoint holding the keys
Axis 1 answers anyone who can reach it. Sitting on the same server are a database password, a directory account and an integration login that still works over legacy authentication.
After: endpoint gone, secrets short-lived
Nobody can reach the SOAP path any more, or it's been replaced. The app's accounts only reach what the app needs, and the identity provider turns away legacy sign-ins.
- Nobody knows which apps carry axis.jar
- The app's database account can read every table
- Integration accounts still sign in to Microsoft 365 with just a password, over legacy protocols nobody switched off
- Directory service accounts have pre-authentication switched off and old passwords
- Someone searches the SBOM for Axis 1 and has an answer in minutes, not a week of emails
- Each app has its own database login with only the rights it needs
- Conditional Access blocks legacy sign-ins for users, and service principals get their own workload identity policy
- Service accounts sit in a vault and get rotated on a schedule. None of them skips Kerberos pre-authentication
- The app server can't open outbound connections to the internet
Several of those changes map to steps in the documented attack, and two deserve a closer look. Microsoft's guidance is to block legacy authentication with a Conditional Access policy, starting in report-only mode so you see what breaks. It also says policies scoped to users don't block service principals, so check those separately. On the directory side, AS-REP roastingdefinitionAn attack on Active Directory accounts that have Kerberos pre-authentication turned off. The attacker asks for a ticket and cracks the password offline. only works against accounts that have Kerberos pre-authentication turned off. Finding those accounts is a quick query, and most of them can have it switched back on.
Our post on the PHP2Shell and Node.js releases walks through scoping an app's own account, and an AI agent is a privileged user covers the same thinking for machine identities. If integration accounts still sign in to Microsoft 365 with passwords, our Entra ID page explains how we move them to modern authentication.
If you've been breached
If you find signs someone used this
Rapid7 published IP addresses and tunnel domains from the cases it saw, and says their absence doesn't prove you're clean. Atlassian says it can't tell whether your instance was hit. Work in order, and keep evidence before you change anything.
- 01Containfirst hour
- Take the server off the network or behind a block-all rule, and snapshot it before restarting anything.
- Turn off the SOAP or legacy API if the server has to stay up for the investigation.
- Copy logs off the box: web server access logs, application logs and the proxy logs in front of it.
- 02Investigatefirst days
- Search access logs for unauthenticated POSTs to the Axis path that got an HTTP 500 back. That's the pattern Rapid7 saw.
- Check whether the app's Java process started bash, curl or other shells, and what it ran.
- Look for outbound connections to throwaway tunnels on Cloudflare Tunnel or Pinggy. Rapid7 saw scripts fetched from both.
- Pull the queries the app's database account ran. Then check where else its stored passwords were used.
- 03Rotatedays 1 to 2
- Change every secret the server held. Start with database passwords, directory bind and Crowd accounts, then API tokens and integration logins.
- If an integration or service account shows up in Microsoft 365 or Entra sign-in logs from an address you don't know, reset it.
- Revoke refresh tokens and sessions for those accounts as well, not just passwords.
- 04Rebuilddays 2 to 5
- Rebuild the server on a fixed version from known-good media rather than cleaning it in place.
- Restore data from before the first suspicious request, then re-apply the mitigations before going live.
- 05Report and learnfirst weeks
- Bring in legal early. Notice duties depend on the data and your state or sector, and Rapid7 saw payment card numbers targeted.
- Call your insurer before hiring outside responders, and ask what the policy needs from you.
- Write down every gap you found. Then put each one on the inventory or the next access review, with a name next to it.
The Rapid7 case is a good map of where to look, because the attacker went for identities early. They queried the Atlassian Crowd API for users and groups, then tried AS-REP roasting against a domain controller, and the roasting attempts failed. A few hours in, they had a service account credential and used it on a production cloud data warehouse. So the investigation can't stop at the Confluence box. Follow every account the box could use.
What this could look like: three typical setups
What this could look like: three typical setups
- the setup
- A firm runs Confluence Data Center 8.5 for its internal wiki. Years ago an admin turned the remote API on for a reporting script, and the server is reachable from the internet for remote staff.
- what happens
- The server answers the crafted SOAP fault without asking for a login. Whoever sends it can read attachments, the Confluence database and every password stored in its configuration.
- what to do
- Turn the remote API off now and check for the 403. Upgrade to 9.2.28 or 10.x, then go through the breach steps, because the API was exposed while attacks were running.
- the setup
- A managed service provider hosts N-central on its own server and turned on the legacy API for an old integration. It manages a few thousand endpoints for its customers.
- what happens
- An attacker who reaches the server can run code on the tool that manages every customer endpoint. Old Agents and Probes can also be impersonated until they're upgraded.
- what to do
- Upgrade to 2026.4 HF3, which blocks the legacy APIs for good, and then upgrade every Agent and Probe. Rebuild the integration on a supported API and tell customers what you checked.
- the setup
- A manufacturer runs a supplier portal bought years ago. Its WAR file carries axis-1.4.jar and a Spring library, and the vendor hasn't said anything about the CVE yet.
- what happens
- Nothing has gone wrong, but the portal fits the exploitable pattern. It has Axis 1 and a known gadget on the same classpath, and the SOAP path faces suppliers.
- what to do
- Block the Axis path at the proxy except for the suppliers' addresses, and ask the vendor in writing for a fix date. If nobody answers, plan a replacement for the integration.
These three are made up, and none is a real incident or a Cohash client. The versions and settings in them come from the Atlassian, N-able and Apache pages behind this post.
How Cohash helps
Most of the damage in this attack came through the accounts the server ran under, and that's the part we work on. We help teams track down old components like Axis 1, close the exposed endpoints, and then vault and rotate the service accounts. Our privileged access management team does the vaulting, our identity threat detection people hunt through sign-in logs, and our DevSecOps practice keeps the SBOMs. If you want someone else to look at your exposure, start with an assessment.
What to take from this
- 01
Every Apache Axis 1 release up to 1.4 has the flaw, and Apache won't ship a fixed 1.x release.
- 02
Confluence Data Center needs 9.2.28 or 10.x, or the remote API turned off. Check that the SOAP path returns 403.
- 03
On-premises N-central needs 2026.4 HF3, and every Agent and Probe needs upgrading too.
- 04
Service authentication doesn't help, since Axis parses the fault before any per-service check runs.
- 05
Search for Axis 1 copies bundled inside WAR files or shaded into fat jars. Axis2 is a separate project and isn't in this CVE.
- 06
After patching, rotate every credential the server stored and block legacy authentication for service accounts.
01Am I affected by CVE-2026-90436?
You are if you run software that includes Apache Axis 1.x, any version up to 1.4, and accepts SOAP requests from people you don't fully trust. The known cases are Confluence Data Center 1.x to 9.x with the remote API on, and on-premises N-central with the legacy API on. Other products that bundle axis.jar may be affected too, so ask their vendors.
02Is there a patch for Apache Axis 1.4?
No release, no. Apache put the code change up as commit 41f3b4e and doesn't expect a 1.x release to follow. You can build Axis yourself with it applied. Apache would rather you moved to a maintained SOAP engine, and it points to Axis2/Java.
03Is Confluence Cloud affected?
No. Atlassian says Cloud sites on atlassian.net or jira.com aren't affected and need no action. Confluence 10.x Data Center isn't affected either, because the remote API was removed in 10.0.
04Is Axis2 affected?
Apache's advisory and the CVE record only list Axis 1.x. Axis2/Java is a separate, maintained project, and it's the replacement Apache recommends. Search carefully, though, since some systems carry both.
05Will turning off the Confluence remote API break anything?
It can. Atlassian warns that scripts, automations and third-party add-ons that use the XML-RPC or SOAP API may stop working. Test it on a non-production site if you can, but don't leave an exposed server waiting on that test.
06We're not a federal agency. Does the November 21 date matter to us?
Legally it only binds US federal civilian agencies. In practice a KEV listing means attacks are already happening, so it's a sensible deadline for anyone else, and federal customers may ask contractors whether they met it.
07We patched. Do we still need to investigate?
Yes, if the API was exposed while attacks were running. Rapid7 saw an attacker go from the first request to stored database passwords within minutes. Check the logs and rotate stored credentials, even if the server looks fine now.
Three questions, about a minute
Pick an answer and we'll tell you if it's right, and why it is or isn't. The score stays in this browser.
-
Q1Your Confluence 9.2 Data Center server has the remote API on, but every SOAP service on it asks for a login. Are you safe from CVE-2026-90436?
show the answer
B. No. Axis parses the fault before any service's login check runs Apache says the fault is read while Axis is still working out the target service, before per-service handlers run. Authentication on the service never gets a say.
-
Q2A dependency scan finds axis2-kernel-2.0.2.jar in one of your apps. What does that tell you about this CVE?
show the answer
B. Nothing on its own. Axis2 is a separate project, but check for an Axis 1 jar as well The CVE covers Apache Axis 1.x only. Axis2 is the separate project Apache recommends moving to, but an app can carry both, so look for Axis 1 classes too.
-
Q3You've switched off the Confluence remote API. What's the quickest way to check it worked?
show the answer
B. Open the soap-axis1/confluenceservice-v2 path and expect a 403 Atlassian says a correctly applied mitigation returns HTTP 403 Forbidden on that path, and no restart is needed.