Skip to main content

CVE-2026-34197: A 13-Year-Old ActiveMQ Bug, Found With AI Assistance

· 8 min read
Ankush K
Co-founder and CTO, Dwaar
CVE-2026-34197: A 13-Year-Old ActiveMQ Bug, Found With AI Assistance

If you run ActiveMQ Classic, your broker is likely one of those systems nobody has looked at since it started working. CVE-2026-34197 is a good reason to look today: anyone who can log in to the web console can run code on the broker, and on some 6.x versions they do not need to log in. Apache published the advisory on April 6, and the fixed releases are already out.

ItemDetail
CVECVE-2026-34197
Affectedactivemq-broker and activemq-all before 5.19.4, and 6.0.0 before 6.2.3
Fixed in5.19.4 and 6.2.3
Severity8.8 High (CVSS 3.1), scored by CISA-ADP on the NVD record. Apache rates it Important. NVD has not published its own analysis
Exploited as of 2026-04-08No public report in any source I have read. Not in CISA KEV
DisclosedApril 6, 2026 on oss-security. NVD published it on April 7

How a management call becomes code execution​

Three pieces combine, and each one looks reasonable alone.

  1. The Jolokia policy. The web console serves Jolokia, a JMX-over-HTTP bridge, at /api/jolokia/. In Apache's words, "The default Jolokia access policy permits exec operations on all ActiveMQ MBeans (org.apache.activemq:*)". That covers BrokerService.addNetworkConnector(String) and addConnector(String), so an authenticated console user can call both over HTTP.
  2. The discovery URI. Those methods take a URI. A crafted discovery URI reaches the VM transport, whose brokerConfig parameter says where an embedded broker gets its configuration. SecurityWeek notes that if the URI names a broker that does not exist, ActiveMQ creates one, and Apache's VM transport reference lists that create behavior as on by default.
  3. The Spring context. brokerConfig is resolved through ResourceXmlApplicationContext, which can fetch XML from a remote URL. Spring instantiates all singleton beans while it builds the context, before BrokerService validates the configuration. A bean definition can invoke a method such as Runtime.exec(), and the code runs in the broker JVM.

The validation does run, just after the damage, so a failed connector add tells you nothing about whether the code executed. SecurityWeek, citing the finders, also describes the flaw as a bypass for CVE-2022-41678.

"Authenticated" depends on what you deployed​

Apache calls this an authenticated bug, and the CVSS vector agrees with PR:L. That holds only for some deployments.

On 6.0.0 through 6.1.1, a separate flaw, CVE-2024-32114, leaves the API web context unauthenticated in the default configuration. The finders' explanation, quoted by SecurityWeek, is blunt: "Jolokia is completely unauthenticated on ActiveMQ versions 6.0.0 through 6.1.1". On those versions, this is remote code execution with no login.

Elsewhere, the login is only as strong as the password. If your console still has the credentials it shipped with, treat it as unauthenticated too. A single base score cannot express any of this, which is the argument for reading CVSS as one input rather than as risk.

What to check and do now​

  1. Find every copy. Check standalone brokers and container images, then search Java dependency trees for activemq-broker and activemq-all, the artifacts Apache lists. For an embedded broker, check whether the console is exposed at all.
  2. Upgrade. On 5.x, go to 5.19.4 or later. On 6.x, go to 6.2.3 or later. Apache lists no other workaround.
  3. Do not be thrown by 5.19.5. The oss-security mail says to upgrade to 5.19.5 while listing 5.19.4 as the affected boundary. Apache's hosted announcement, the NVD description, SecurityWeek and BleepingComputer say 5.19.4, which is where the fix landed. 5.19.5 is a real maintenance release dated April 8, and its changelog does not mention this CVE. Either version covers this bug, so take the newer one if you can.
  4. Do not stop at 6.1.2. It fixes CVE-2024-32114 by enabling authentication by default, but it is still inside the affected range for this CVE. The authentication constraint in conf/jetty.xml that NVD gives for CVE-2024-32114 closes the no-login path only.
  5. Search your broker logs. BleepingComputer reports the indicator as connections using the VM transport with brokerConfig=xbean:http://, and says a configuration warning can mean the payload already ran. If you find one, I would treat that broker as compromised.
  6. Shrink the surface. This is my advice, not Apache's. Take the console and /api/jolokia/ off every network that does not need them, change default credentials, and restrict outbound traffic from broker hosts, since the published attack has the broker fetch remote XML.

Look at the dates as well. BleepingComputer reports that the bug reached Apache on March 22, and the fixed releases shipped in late March. The advisory came about a week later. If your patching starts when an advisory lands, you started behind.

Old code is being re-read​

Apache credits Naveen Sunkavally as the finder. BleepingComputer reports that he found it using Claude with a couple of basic prompts, with the result described as "80% Claude with 20% gift-wrapping by a human". The 13-year figure also comes from the finders, and SecurityWeek and BleepingComputer repeat it. Apache's advisory mentions neither the age nor the AI, so treat both as the finder's account.

One bug is not a trend. My position is narrower: the cost of reading old code just dropped, and that changes what age means.

This bug requires holding three things in your head: an access policy, a transport URI parameter and the order Spring does things in. A careful reviewer could connect those, but a stable old broker rarely gets a reviewer with that much patience. A long history without this kind of finding used to be weak evidence of safety. I now treat it as close to no evidence, because ActiveMQ has had plenty of attention and this path across components was still missed.

No source I have read says attackers found this bug with a model or are exploiting it. I am assuming anyone can ask the same questions of the same public code, and planning on that basis.

The window after disclosure is already short. The unauthenticated Langflow bug from March came under active exploitation within 20 hours of public disclosure, according to The Hacker News. Here, the mechanism has been public since April 6, and the finders' own write-up was in the trade press by April 8. No reported exploitation is a statement about today, not a reason to put this upgrade behind other work.

So stop ranking middleware by how recently it had a CVE. Know where it runs, know who can reach its admin API, and be able to upgrade it this week.

Pick the oldest piece of middleware you run that has an admin API. Write down its version, who can reach that API, and what its default policy allows. Then have a coding model trace every path from that API to class loading or process execution, and check what it gives you by hand. If the answer surprises you, you have your next upgrade ticket.

References​