Two supply-chain breaks in one day: the JVM and the agent framework
Java4Shell allows pre-authentication remote code execution in OpenJDK, while two malicious agent-framework releases on PyPI stole cloud credentials on every Python start.
- Java4Shell (CVE-2026-114938) is a heap overflow in OpenJDK's AVX-512 Base64 decoder; Calif rates it CVSS 4.0 9.2 while Oracle rates it 8.1. m:…9552
- Oracle fixed builds for Java4Shell include 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1, and the flaw can be mitigated by starting the JVM with -XX:-UseBASE64Intrinsics. m:…0628
- Calif demonstrated pre-authentication code execution in default configurations of Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate. m:…9552
- Microsoft's agent-framework releases 1.26.1 and 1.26.2 on PyPI contained a credential stealer that runs on every Python start through a .pth file. m:…6615
- Both malicious agent-framework versions have been removed from PyPI, and users should pin agent-framework to version 1.26.0. m:…8000
- Version 1.26.2 added logic to steal GitHub Actions OIDC tokens and was designed to evade diff-based detection by keeping the main module byte-identical to the clean version. m:…8000 m:…2873
A flaw in the runtime itself
Java4Shell, tracked as CVE-2026-114938, is a heap overflow in OpenJDK's AVX-512 Base64 decoder. A malformed padded tail makes the decoder write 62 bytes past a 50-byte output array, and the flaw affects OpenJDK 17, 21, 25, 26 and 27 on x86-64 machines with AVX-512 support. ARM64 architectures, including AWS Graviton, are not affected. The vulnerability was disclosed by Calif on November 17, 2026, and Oracle released patched builds the same day. m:…9552 m:…0628
The reason this flaw matters is how easily it is reached. A remote attacker with no credentials can overwrite heap memory using the vulnerability, because standard authentication handling passes raw tokens to the Base64 decoder before validation, and one widely used JWT library decodes JWT headers and payloads before checking the signature. The trigger is a string of 64 A characters followed by an equals sign. Calif confirmed remote code execution in Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate without authorization. No exploitation in the wild had been reported as of November 18, 2026, but the exploit itself was built fast: OpenAI Codex helped find the bug in one hour and build an exploit in two hours. m:…2672 m:…9552
There is also an uncomfortable detail in the history. A fix for the bug was proposed in a public pull request on September 9, 2026, but was auto-closed without merging, leaving the vulnerability present in shipped builds for months. Chainguard discovered the bug while scanning an open source project, and has published fixed packages for OpenJDK 8, 11, 17, 21, 25 and 27. The immediate response is to patch to the fixed builds, including 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1, or to start the JVM with the flag that disables the vulnerable intrinsic. Java4Shell is a runtime-level vulnerability, unlike Log4Shell, which was a library-level flaw, so the exposure sits underneath every Java application on an affected machine rather than inside one dependency. m:…6616 m:…0628
A package you already trusted
The second break came through the package index rather than the runtime. Microsoft's agent-framework package releases 1.26.1 and 1.26.2 on PyPI contained a credential stealer, and the compromise was an account takeover of the Microsoft-associated PyPI publisher, not a typosquatting attack. The malware uses a .pth file in site-packages to execute code on every Python interpreter start, so it runs on pip, pytest and IDE language servers, not just on code that imports the package. It exfiltrates environment variables matching patterns like TOKEN, SECRET, PASSWORD, KEY, AZURE, AWS, GITHUB and OPENAI, along with GitHub Actions OIDC tokens, Azure DevOps OIDC tokens and Azure managed identity tokens. m:…6615
Two malicious releases hit PyPI 33 minutes apart, and the second one expanded the attack: version 1.26.2 added specific logic to steal GitHub Actions OIDC tokens. The stealer targets Azure, Azure DevOps and GitHub Actions secrets, and exfiltrates data to a webhook.site URL. The package sees approximately 190,000 downloads a month. The malicious releases do not fetch or execute remote code, and they pin the core framework to version 1.26.0 so the rest of the framework installs normally and applications keep running while the stealer operates in the background. m:…8000 m:…2873
Detection was fast but the design was deliberate. One scanner blocked agent-framework 1.26.1 at 20:36:08 UTC, 48 seconds after publication, and blocked 1.26.2 at 21:09:21 UTC, 42 seconds after publication. Version 1.26.2 was designed to evade diff-based detection by keeping the main module byte-identical to the clean version, and the malicious releases do not ship a source distribution, unlike previous legitimate releases. The GitHub source for agent-framework is not reported as modified, which points to the publisher account rather than the code repository as the point of entry. m:…6615 m:…2873
Two different breaks, one pattern
The two incidents are different in kind and alike in effect. One is a bug in shared infrastructure that nobody chose and everybody inherits: a runtime flaw that turns an ordinary authentication header into a path to code execution. The other is a trusted publisher account turned against its own users, with a payload that runs before any application code and steals the credentials that protect cloud and CI systems. In both cases the dangerous thing arrives through a channel developers do not inspect line by line: the JVM they run on and the package index they install from. m:…0628 m:…6615 m:…9552
The fixes are concrete and unglamorous. Patch Java to the fixed builds or disable the vulnerable intrinsic, and pin agent-framework to version 1.26.0 while checking whether the malicious versions were installed anywhere. No exploitation of Java4Shell in the wild had been reported as of November 18, 2026, and both malicious agent-framework versions have been removed from PyPI. The open question is how many machines and pipelines picked up either problem in the window before the fixes landed, and whether the next break will again arrive through the layers everyone assumes are already safe. m:…0628 m:…8000 m:…9552
What is still unresolved
- How many systems installed agent-framework 1.26.1 or 1.26.2 during the window they were live on PyPI?
- How many Java services on affected hardware are still running unpatched builds?
- What will the PyPI publisher-account takeover change about how package publishers protect their accounts?
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.