Blog Application Security

The Attack Surface You Do Not Inventory: Runtimes, Meta-Packages and Third-Party Scripts

Today's briefing covers a JVM Base64 flaw with pre-auth reach, a poisoned Microsoft AI-agent package on PyPI, and malicious JavaScript served through Optimizely. The common thread is code most teams never list as an asset, and AI speeding up both sides of the exploit race.

Published 5 min read
Line-art map of connected systems with three unlabelled nodes highlighted, representing uninventoried attack surface

Key takeaways

  • A HotSpot JVM Base64 flaw, CVE-2026-114938, reaches Java apps before authentication; Calif.io reports pre-auth code execution on several widely deployed products.
  • Malicious agent-framework releases on PyPI ran a credential stealer at interpreter start via a .pth file, according to Wiz.
  • Optimizely confirmed attackers exploited a vulnerability to place malicious JavaScript in customer projects, which its trusted CDN then served.
  • CISA added a Puppet deserialization flaw to its Known Exploited Vulnerabilities catalog; autosign configuration decides how exposed a server is.
  • Google says it has observed attackers fully automating patch diffing, exploit development and execution, which makes patch latency a primary control.

Three of today's lead stories hit code that rarely appears on an asset list: a hardware-accelerated Base64 path inside the JVM, a meta-package that executes at Python interpreter start, and the custom JavaScript field of an A/B testing tool served from a trusted CDN. Two more stories explain why that gap is getting more expensive: researchers and attackers are both using AI to find and weaponize bugs, and Google says it has now seen the whole loop automated.

Java4Shell: the runtime is the vulnerability #

Calif.io disclosed CVE-2026-114938, a memory-corruption flaw in the AVX-512 VBMI Base64 decoder path of OpenJDK's HotSpot JVM. Oracle fixed it in its latest Critical Security Patch Update, OpenJDK published a matching advisory, and CISA issued a standalone alert urging immediate patching.

Because the bug is in the runtime, it reaches any Java application that decodes Base64 before authentication. The Hacker News cites HTTP Basic login headers and SAML artifacts as examples of such input. Calif reports pre-auth code execution on default installs of Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate. That list comes from Calif's own testing and it says the list is not exhaustive.

Two things help with triage. First, Calif says the flaw needs an x86-64 CPU with AVX-512 VBMI, and that ARM64 is not affected. Second, Calif documents a workaround: launching the JVM with -XX:-UseBASE64Intrinsics. Oracle rates the flaw High and difficult to exploit, while Calif scores it Critical under the newer CVSS standard. CISA has not said it is exploited in the wild.

The real work is inventory. CISA's alert calls out JDKs bundled in third-party applications, container images and appliances. If your JDK list only covers what your own builds install, it is incomplete.

A poisoned AI-agent package that runs without import #

Two malicious releases of Microsoft's agent-framework meta-package were published to PyPI and quarantined the same day. According to Wiz, each dropped a .pth file into site-packages, so the credential stealer ran at every Python interpreter start whether or not the package was imported.

Wiz reports the stealer went after secret-like environment variables, credential files, Azure managed-identity and Azure DevOps tokens, and in the second release GitHub Actions OIDC tokens. Wiz found no matching tag in the project's GitHub repository and notes the first malicious release was wheel-only, but says the root cause is unconfirmed.

The takeaway for CI/CD owners: pin and hash-verify AI framework dependencies like any other, and remember that an OIDC token stolen from a runner is worth whatever cloud role it federates to.

Optimizely: trusted origin, hostile content #

Optimizely confirmed that an unauthorised party exploited a vulnerability in its Web Experimentation product to add malicious JavaScript to some customer projects, and says the flaw is fixed and the code removed. Report URI, which tracked the incident, says the attacker wrote into the Project JavaScript field and that this was not CDN cache poisoning. The trusted CDN simply served what the project contained.

Report URI counted 188 infected snippets out of 1,611 examined, across 178 organisations. Those are its figures, and Optimizely has given none. The payload replaced cryptocurrency addresses, skimmed card data, stole cookies and screenshots, and loaded iOS exploit kits, though SecurityWeek reports Hudson Rock said successful iOS exploitation has not been demonstrated.

A script tag pointing at a vendor with a free-text code field is a write path into your production pages. Treat the vendor console as privileged infrastructure, and use CSP reporting.

Also on the radar: Puppet in KEV #

CISA added CVE-2026-115731, a deserialization flaw in Puppet's report-processing code, to the Known Exploited Vulnerabilities catalog with a two-week federal remediation deadline. Per the CVE record, any client holding a CA-signed agent certificate can run code on the Puppet Server as its service user by submitting a crafted report. Where the CA autosigns requests without strong verification, an unauthenticated network attacker can obtain such a certificate. Perforce has published fixed builds, and the open-source OpenVox fork shipped fixes as well. CISA's entry does not describe the exploitation activity.

Puppet Server is the root of trust for fleet configuration, and Neodyme, which reported the bug, describes the path from code execution to CA key theft. Check your autosign configuration first: it decides whether this is post-auth or effectively unauthenticated.

The through-line: AI on both sides #

Calif says it used an AI coding agent to find and weaponize the Java bug in a matter of hours. Neodyme says AI-based analysis, with strict steering, found both Puppet bugs. Those are researcher statements, but they point the same way.

On the offensive side, Google Threat Intelligence Group reports that in the past quarter it identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop, which it says is a first. Its main example is a multi-agent framework found on an exposed staging server. Google says the framework compares web application code before and after patch announcements, identifies potential vulnerabilities, drafts exploits and executes them automatically. Google also says the operator was interested in French media, entertainment and education targets. The post names no victims, no specific CVEs and no specific models.

If Google's assessment holds, the gap between an upstream fix and your deployment is itself attack surface, and a quiet fix commit is a disclosure.

What to do this week #

  • Build a JDK inventory that includes bundled runtimes in containers, appliances and vendor products, and record CPU architecture alongside it.
  • Search lockfiles, images and CI logs for the quarantined agent-framework releases, and rotate credentials reachable from affected runners.
  • List every third-party script with a free-text JavaScript field, lock down the consoles behind them, and turn on CSP reporting.
  • Audit Puppet CA autosign settings and patch Puppet Server.
  • Measure your time from upstream patch to production for internet-facing web frameworks and plugins, then shorten it.

Sources

  1. Java4Shell: Write once, RCE everywhere, Calif.io
  2. Critical Java Vulnerability Requires Immediate Patching, CISA
  3. Oracle Critical Security Patch Update Advisory - November 2026, Oracle
  4. Microsoft agent-framework Compromised: Credential Stealer Targets Azure and CI Pipelines, Wiz
  5. agent-framework · PyPI (release history), PyPI
  6. Web Experimentation Security Incident, Optimizely
  7. World's Largest Brands Infected With Malware Via Optimizely, Report URI
  8. Optimizely Vulnerability Exploited to Serve Malware on Major Websites, SecurityWeek
  9. CISA Adds One Known Exploited Vulnerability to Catalog, CISA
  10. CVE-2026-115731 Detail, NVD
  11. Pulling the right strings for RCE on a Puppet, Neodyme
  12. Critical CVE Updates in all known versions of Puppet and OpenVox, Vox Pupuli (OpenVox)
  13. GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold, Google Cloud / Google Threat Intelligence Group
  14. Java4Shell Flaw (CVE-2026-114938) Enables Pre-Auth RCE in Jira, Jenkins, Keycloak and More, The Hacker News
  15. OpenJDK Vulnerability Advisory: 2026/11/17, OpenJDK
  16. Known Exploited Vulnerabilities Catalog, CISA

Questions this raises

What connects today's stories?
Each one hits a component that is rarely inventoried as an asset: a JVM intrinsic, a Python meta-package that executes at interpreter start, a third-party script field, and a configuration management server. Researchers in two of the stories say AI helped find the bugs, and Google reports attackers automating exploitation.
Which of these is confirmed as exploited?
The Puppet flaw is in CISA's Known Exploited Vulnerabilities catalog, and Optimizely confirmed its vulnerability was exploited. The agent-framework releases were malicious by design. CISA has not said the Java flaw is being exploited in the wild.
What is the single most useful action for an appsec team?
Close the inventory gap. Know every JDK you run including bundled ones, every pinned AI framework dependency, and every third-party script that can carry custom code, then shorten the time between an upstream fix and your deployment.

daily-briefingsupply-chainAI Security