Blog Vulnerabilities & Zero-Days

The Attack Surface Under Your Code: Axis, LiteLLM, Java, NGINX and a Poisoned PyPI Package

This week's serious issues all sit in layers that appsec teams did not write: an end-of-life SOAP library, an LLM gateway, the Java runtime, the NGINX front door and a trusted PyPI package. Here is what happened in each, and the inventory, patching and rotation work that follows.

Published 6 min read
Line art of stacked layers beneath an application window with cracks in the lower layers, beside the briefing headline

Key takeaways

  • This week's code execution and credential theft issues sit in runtimes, embedded libraries, gateways and packages, not in application code.
  • An end-of-life Apache Axis library embedded in Confluence Data Center and N-able N-central is under active exploitation, per Rapid7 and CISA.
  • Java4Shell puts a JDK Base64 bug in front of authentication. Code execution is a Calif claim and no exploitation is reported.
  • Malicious agent-framework releases on PyPI ran a credential stealer on every Python start.
  • Start with inventory, then patch or apply workarounds, then rotate credentials where a malicious package may have executed.

None of this week's lead stories involve a bug in code your team wrote. Every one of them lands inside applications your team is responsible for. An end-of-life SOAP library, an LLM gateway, the Java runtime, the NGINX front door and a trusted PyPI package each produced code execution or credential theft within days of each other.

The through-line is simple: the layers underneath the application are the attack surface.

An end-of-life SOAP library, still shipping #

CISA added four flaws to its Known Exploited Vulnerabilities catalog, including CVE-2026-90436 in Apache Axis. Rapid7 describes it as an unsafe reflection flaw in the original, legacy Axis line that allows unauthenticated remote code execution. That line is end-of-life, and Rapid7 says Apache does not plan a fixed release for it.

The library lives on inside commercial products. Rapid7 names Atlassian Confluence Data Center and N-able N-central, both vulnerable only in a non-default configuration. For Confluence, that is the 'Remote API (XML-RPC & SOAP)' setting. Rapid7's guidance is to upgrade Confluence to a fixed release or turn that setting off, and it says N-able has shipped an N-central hotfix that disables the 'Legacy API' feature.

Rapid7 says it detected exploitation against Confluence Data Center in a customer environment. The attacker used Confluence's own database credentials to read page content and attachments, searched for credentials and payment card numbers, queried a cloud data warehouse with a service account about four hours after the first exploit request, and signed in to Microsoft 365 over legacy authentication about twelve hours after it. A wiki compromise became a cloud data compromise inside one working day.

The AI gateway that holds every key #

The same KEV update lists CVE-2026-40217 in BerriAI LiteLLM. Google's Threat Intelligence Group reports a threat actor exploited it in LLMjacking intrusions, and describes it as a sandbox escape that elevates limited-access inference keys, called Virtual Keys, into full administrative credentials.

A gateway in front of model APIs concentrates trust. Whoever owns its admin credential owns every provider key behind it. Treat it like any other secrets broker: restrict who can reach it and watch key activity.

A runtime bug in front of the login check #

Research firm Calif disclosed CVE-2026-114938, which it calls Java4Shell: a heap overflow in the OpenJDK Base64 decoder intrinsic that uses AVX-512 VBMI instructions. Oracle has shipped fixes. The Canadian Centre for Cyber Security says the flaw is remotely exploitable without authentication by sending crafted Base64 data to a Java process on an affected x86-64 processor.

Java servers often decode Basic auth headers, tokens and SAML artifacts before authenticating anyone, The Hacker News notes, so the vulnerable path is reachable in applications with no flaw of their own. Calif claims pre-authentication code execution in default configurations of Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate, and is withholding those exploit chains. That is a researcher claim. The Hacker News reports other labs reproduced crashes and a memory leak but not code execution, and the Cyber Centre says it is not aware of active exploitation.

Workarounds exist where patching has to wait. -XX:UseAVX=2 disables AVX-512 in the JVM, and Calif also lists -XX:-UseBASE64Intrinsics. Calif warns that the first flag must only be used on x86-64, because OpenJDK on ARM will fail to start with it.

A trusted package turned stealer #

Two malicious releases of Microsoft's agent-framework package reached PyPI and were removed within the hour, according to Elastic Security Labs and Wiz. Both launched a credential stealer from a .pth file, so it ran on every Python start in the environment, whether or not the package was imported.

The stealer went after secret-like environment variables, credential files, Azure CLI token caches, Azure managed identity tokens and, in the second release, GitHub Actions OIDC tokens. Wiz's analysis is that a client certificate and key for Microsoft's ESRP publishing service appear to have been compromised, with the method of theft still unknown.

Both vendors say to pin to the last clean release, search site-packages for agent_framework_meta.pth and af_runtime_diag.pth, and rotate every secret readable on an affected host.

The front door and an old blog engine #

Two more items fit the pattern.

NGINX. The nginx.org security advisories page lists five memory-safety flaws fixed in the latest mainline and stable releases. The lead one is CVE-2026-123007, a memory corruption in HTTP/2 flow control that nginx.org rates as major. The other four are use-after-free bugs in the mail proxy module, HTTP/2, HTTP/3 and the resolver. Press coverage is more alarmed than the nginx.org ratings: The Hacker News reports that researchers at Wiz and Xint demonstrated code execution through the HTTP/2 flaw in testing and that a proof of concept is public, while no exploitation in the wild has been reported. The same report, citing CIRCL guidance, says disabling HTTP/2 is a temporary mitigation that blocks only the two HTTP/2 bugs. The Hacker News also advises a full restart or live binary upgrade after upgrading, because a reload leaves the old executable running. Check the real binary in each container image.

Apache Roller. The Apache Software Foundation published an advisory for CVE-2026-129448, a critical unauthenticated deserialization flaw in Roller's XML-RPC endpoint, and states plainly: 'This issue is being actively exploited.' The advisory says the endpoint deserializes vendor extension types before authenticating the requester, and that it is the same issue as the earlier CVE-2026-82384. SecurityWeek reports the earlier advisory listed a narrower affected range, so anyone who triaged it as not applicable should look again. Turning off XML-RPC in the admin UI is not sufficient, per Apache, because older releases deserialize the request before checking that setting. The workarounds are to set enabledForExtensions to false in WEB-INF/web.xml, or remove the /roller-services/xmlrpc servlet mapping, then restart.

The clock is shorter #

Google's threat intelligence team also reports an attacker framework that diffs patches and launches exploits with no human in the loop. If that holds up, the quiet period defenders assume after a fix ships gets shorter.

What to do this week #

  1. Inventory first. Which JDK builds, embedded libraries, NGINX binaries and pinned package versions are actually deployed, including inside vendor appliances and container images?
  2. Patch or apply the documented workaround. Fixes or documented workarounds are cited above for the Axis exposure in Confluence and N-central, the JDK, NGINX and Roller. For LiteLLM, the sources here give no fix detail, so restrict access to the proxy and check the vendor advisory.
  3. Turn off legacy endpoints. The Confluence Remote API, the N-central Legacy API and Roller's XML-RPC mapping are all remote interfaces from an older era. Rapid7 notes that integrations using the Confluence API may stop working, so check for dependants first.
  4. Rotate where a malicious package may have run. Any host or pipeline that resolved an unpinned agent-framework during the exposure window is a credential compromise until shown otherwise.
  5. Pin dependencies. A pinned build would not have pulled the poisoned releases.

Sources

  1. CISA Adds Four Known Exploited Vulnerabilities to Catalog, CISA
  2. Rapid7 Detected Exploitation of CVE-2026-90436 against Atlassian Confluence Data Center, N-able N-central also Exploited, Rapid7
  3. GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold, Google Cloud (Google Threat Intelligence Group)
  4. Java4Shell: Write once, RCE everywhere, Calif
  5. AL26-029 - Critical vulnerability impacting Oracle Java - CVE-2026-114938 - Canadian Centre for Cyber Security, Canadian Centre for Cyber Security
  6. Java4Shell Flaw (CVE-2026-114938) Enables Pre-Auth RCE in Jira, Jenkins, Keycloak and More, The Hacker News
  7. Microsoft agent-framework Compromised: Credential Stealer Targets Azure and CI Pipelines, Wiz
  8. We caught the agent-framework PyPI compromise six minutes after it published, Elastic Security Labs
  9. nginx security advisories, nginx.org
  10. 10-Year-Old NGINX HTTP/2 Flaw Lets Unauthenticated Attackers Execute Code on Servers, The Hacker News
  11. CVE-2026-129448: Apache Roller: Unauthenticated deserialization in the XML-RPC endpoint (releases before 6.1.5), Apache Software Foundation
  12. Critical Apache Roller, Axis Vulnerabilities Exploited in Attacks, SecurityWeek

Questions this raises

Which of these issues are confirmed as exploited?
CISA lists the Apache Axis and LiteLLM flaws in its Known Exploited Vulnerabilities catalog, and Apache's advisory states the Roller flaw is being actively exploited. The agent-framework releases were malicious by design. No exploitation in the wild has been reported for the Java Base64 flaw or the NGINX flaws.
Where should a team start?
With inventory. Each issue depends on what is actually deployed: the JDK build and CPU, embedded copies of the legacy Axis library, the NGINX binary inside each image, and the package versions your builds resolved.
Does disabling XML-RPC in the Apache Roller admin UI mitigate the flaw?
No. The Apache advisory says that is not sufficient, because older releases deserialize the request before checking that setting. Upgrade, or apply the web.xml workarounds in the advisory and restart.

daily-briefingsupply-chainvulnerabilities