Calif Research said on November 17 that one of its researchers used AI to find Java4Shell, a memory-corruption flaw in OpenJDK that CISA has called critical and urged organizations to patch immediately.

Java runtimes are often bundled inside third-party applications, containers and appliances, and CISA warned that they may not appear as separate entries in software inventories1. Calif warned that the vulnerable decoder can be reached through everyday functions2 such as authentication and file uploads. Endor Labs said the fix was not yet in some public OpenJDK update repositories as of November 17, and Cisco Talos had not yet seen exploitation in the wild.

The flaw, tracked as CVE-2026-114938, affects the Base64 decoder of the x86-64 HotSpot Java Virtual Machine. According to CISA, specially crafted Base64 input could crash the JVM and allow unauthenticated remote code execution.

Calif's findings and its account of AI's role reach this report through SecurityWeek. Calif's own write-up and Oracle's advisory were not reviewed. CISA's alert does not mention Calif or AI.

How Calif says AI found the flaw

Calif said AI played a key role in both finding and exploiting the bug. "It took our own Quang Luong an hour to discover the vulnerability, and two more hours to develop the first working exploit. The total token cost was less than $200," the company said, as quoted by SecurityWeek.

Neither SecurityWeek's report nor CISA's alert names the AI tool used. A post on X by @loclangi3 attributes the work to OpenAI's Codex. The post is a third-party summary with no independent confirmation.

Calif said that in initial tests it achieved pre-authentication remote code execution on default configurations of Keycloak, Atlassian Jira and Confluence, JFrog Artifactory, Jenkins and PingFederate. It also reported authenticated code execution on Apache Kafka Broker, Solr, Elasticsearch, Gerrit and Spinnaker, and said neither list is exhaustive.

Calif also described a memory-leak capability it calls JavaBleed, which it said could expose PII, secret keys and tokens from the heap. A separate exploit shared with Calif by security researcher samczsun, it said, can compromise a Java application that connects to an attacker-controlled server over TLS.

Calif said there is no universal exploit, because the steps depend on how an application uses Base64. Its public proof-of-concept code demonstrates memory corruption but not code execution. The company said weaponized chains for individual products will stay private until others have had a chance to upgrade.

VulnCheck gave a more cautious assessment. It said crashing a vulnerable JVM is easy once an attacker finds a reachable entry point, and that it has built a full remote code execution exploit itself. But it said achieving that is much harder, because the attacker must manipulate the heap layout and bypass ASLR in a way specific to each target. "We don't anticipate that CVE-2026-114938 will be easily or widely exploitable," VulnCheck said, adding that adversaries are likely to find narrow paths in specific applications.

Calif assigned the flaw a CVSS 4.0 score of 9.2. Oracle's advisory, as described by SecurityWeek, gives it a CVSS 3.1 score of 8.1 and lists it as remotely exploitable without authentication but with high attack complexity.

Patches and affected systems

SecurityWeek reported that news of the flaw surfaced ahead of the coordinated disclosure. Oracle released fixes in its November 2026 Critical Security Patch Update. Endor Labs said users of other Java distributions must wait for vendors to roll out patched builds, and that as of November 17 the fix was not yet in the public OpenJDK 17, 21 and 25 update repositories.

According to Calif, OpenJDK versions starting with 17.0.6 are vulnerable, as are distributions based on it, including Amazon Corretto, Eclipse Temurin, Microsoft Build of OpenJDK, Azul Zulu, Red Hat OpenJDK and Oracle Java SE. The OpenJDK advisory lists Java 17, 21, 25 and 27 as affected.

Exploitation requires x86-64 hardware with AVX-512 VBMI support, such as Intel Xeon processors from the Ice Lake generation onward and AMD chips based on Zen 4 or newer. SecurityWeek reported that ARM64 systems are not vulnerable.

CISA said organizations that cannot patch immediately should use compensating controls: restricting and validating Base64 input at the application boundary, enforcing input-size limits, and limiting unauthenticated access. It separately advises monitoring for rogue runtimes, unexpected JVM crashes and suspicious Java processes, and prioritizing internet-facing systems for patching.

Cisco Talos has yet to see the vulnerability exploited in the wild, according to SecurityWeek, but said it anticipates "probing attempts may pick up soon after public disclosure."