Java4Shell (CVE-2026-114938) is a heap overflow in the Java virtual machine's fast Base64 decoder. It can let an unauthenticated attacker run code inside common Java servers. It only fires on x86-64 processors with a specific vector extension, AVX-512 VBMI. Most Intel Xeon chips since Ice Lake and AMD EPYC Zen 4 parts have it, so a large share of the rented servers running Java today do. If you run Java 17 or newer on one of those chips and your package manager has not delivered a fix yet, the JVM startup flag -XX:-UseBASE64Intrinsics is the control you can apply this morning.
Are you affected? Three checks per host
Exposure is decided host by host, by three facts together: the Java version, the CPU features the JVM sees at startup, and whether the Base64 fast path is switched on. All three have to line up, which is why two servers running the same Java package can land on opposite sides of this bug.
-
Version. any runtime derived from OpenJDK at 17.0.6 or newer, which takes in the 21, 25 and 27 lines too, according to the Calif research team that found it. Standard OpenJDK 8 and 11 are not affected. Oracle's Java SE 8 Enterprise Performance Pack build 8u511b33-perf is, per the CVE record.
-
Processor. The CPU must report all four of
avx512f,avx512bw,avx512vbmiandbmi2. ARM64 servers are out of scope. -
Intrinsic enabled. HotSpot turns the accelerated decoder on by default when the hardware supports it, so on an affected chip you are exposed unless someone has switched it off.
One nuance the advisories skip: the JVM reads processor features through the CPUID instruction inside whatever machine it runs on. On a virtual server, the guest's view of the CPU is what counts, not the host's spec sheet. A cloud instance or VPS that is live-migrated or resized onto newer hardware can pick up AVX-512 VBMI. A Java service that was not exposed last week then becomes exposed on its next restart, with no package change at all. So check from inside each guest, and re-check after any instance-type change.
grep -o -w -E 'avx512f|avx512bw|avx512vbmi|bmi2' /proc/cpuinfo | sort -u java -version java -XX:+PrintFlagsFinal -version | grep -E 'UseBASE64Intrinsics|UseAVX ' find / -xdev -name 'hs_err_pid*.log' -newermt '2026-11-17' 2>/dev/null export JAVA_TOOL_OPTIONS="-XX:-UseBASE64Intrinsics"
Read it top to bottom. If the first line prints fewer than four flags, that host is outside the bug's reach. The third line shows what the running JVM actually decided, which is the only answer that matters for a product shipping its own runtime. The find looks for HotSpot fatal-error logs written since disclosure. A burst of them from an internet-facing Java service is the cheapest early signal you have. The last line is one way to push the workaround into every JVM a service starts; a service-manager or product-specific setting works too.
Why the distro package is the bottleneck
Oracle shipped fixed builds on November 17, 2026, the same day the OpenJDK Vulnerability Group advisory went out: 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1, plus 8u513-perf for the Performance Pack, per Qualys's write-up. Amazon Corretto had a fixed 17.0.21.1.2 build the same day.
| Java line | Status | Oracle fixed build |
|---|---|---|
| 8 (standard OpenJDK) | Not affected | n/a |
| 8 Performance Pack | 8u511b33-perf affected | 8u513-perf |
| 11 | Not affected | n/a |
| 17 | 17.0.6 and later affected | 17.0.21.1 |
| 21 | Affected | 21.0.13.1 |
| 25 | Affected | 25.0.5.1 |
| 27 | Affected | 27.0.1.1 |
Most Linux servers do not run Oracle's binaries. They run openjdk-17 or java-21-openjdk from the distribution, and those maintainers rebuild on their own schedule. When we checked the Debian security tracker early on November 18, every OpenJDK source package from 17 to 27 was marked vulnerable with no fixed version, including the bookworm and trixie security builds. Ubuntu had not yet published a page for the CVE. Endor Labs separately reported that the fix had not yet reached the public jdk17u, jdk21u and jdk25u update repositories when the advisory landed.
So for an apt or dnf shop, running upgrade today will most likely leave you on a vulnerable build that looks current. That gap is the real exposure window, and the flag is what covers it. Watch your distribution's tracker and switch back to the stock decoder once a fixed package is installed.
Which servers to flag first
Calif reports pre-authentication code execution against default installs of Keycloak, Jira, Confluence, JFrog Artifactory, Jenkins and PingFederate, and authenticated paths in Kafka, Solr, Elasticsearch, Gerrit and Spinnaker. Login forms, token parsing and file uploads all decode Base64 as a matter of routine, which is why the reachable surface is so wide. On a typical self-hosted stack that means the identity provider and the CI server come first. They sit on the internet more often than a broker does, and they hold the keys to everything else. Our earlier coverage of Keycloak's vault secret leak and Artifactory token forgery shows how far one foothold on those boxes reaches. Confluence admins are already mid-patch from the exploited Axis flaw we covered yesterday. Check whether that upgrade's bundled runtime is a fixed one.
Do not overlook clients. Calif says Java code acting as a TLS client can hit the decoder too, simply by connecting out to a hostile server, so batch jobs and agents that call out are in scope even with no listening port.
Products that bundle their own JRE will not pick up a distro update at all. CISA's alert makes the same point: runtimes inside appliances and container images often never show up in a software inventory. That makes the PrintFlagsFinal check against the product's own java binary more trustworthy than any package list.
Exploitation status: no attacks seen, technique public
As of November 17, neither Qualys nor Cisco Talos had seen exploitation in the wild, and Talos expects probing to start soon. Calif is holding back its product-specific exploit chains for now but has published the technique. Several GitHub repositories carrying the CVE ID appeared within hours, mostly scanners and crash demonstrations of unverified quality. Read any "runtime patch" script before running it on a production box; at least one asks you to pipe a remote script straight into a shell. Talos has shipped Snort rules (SIDs 1:67502 through 1:67513) that match malformed Base64 in HTTP fields.
Severity depends on who is scoring. Oracle rates it CVSS 3.1 8.1, High, with high attack complexity. Calif and Talos rate it 9.2 under CVSS 4.0. The complexity rating reflects that each application needs its own chain. It does not make the bug safe to wait out. This is the second Calif disclosure in a month to get a "2Shell"-style name, after PHP2Shell. That one went from leak to patch to scanning quickly, which is a fair model for what to expect here.
Set the flag tonight, patch when your package lands
On every x86-64 host that passes the CPU check, add -XX:-UseBASE64Intrinsics to the JVM options and restart. Then confirm with PrintFlagsFinal that it took. The flag only switches Base64 to the slower generic code, so test anything that decodes large blobs. The broader -XX:UseAVX=2 also works, but do not ship it in a shared config that ARM hosts read, because the JVM there will refuse to start. Neither replaces the update. Once your vendor or distribution publishes a fixed build, install it, restart, and remove the flag. Keep watching for hs_err_pid files and for Java processes spawning shells; if you track this kind of exposure fleet-wide, our vulnerability detection page covers how we approach it.