You ran apt upgrade openjdk-17-jdk, java -version printed a new build, and you went to lunch. Your Keycloak container still runs its own JRE, your Jenkins agent image is pinned to last month's Temurin, and the vendor appliance in the corner has a JVM you have never seen a version string for.
Java4Shell (CVE-2026-114938) is a heap overflow in HotSpot's AVX-512 Base64 decoder. Calif, who found it, rates it 9.2 under CVSS 4.0 and lists Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate as pre-auth RCE in default configs. Oracle scores it 8.1 under CVSS 3.1, citing high attack complexity (Aikido). Patches shipped November 17, 2026.
Below: which Java 17, 21, 25 and 27 builds are affected, which are fixed, and four commands that find every JVM on a host or in a container image, including the ones java -version never sees.
Is it really the next Log4Shell?
Not in how easy it is to exploit. Log4Shell was one string in a User-Agent header. Here the overwritten bytes are a fixed pattern derived from the decoder, not attacker-chosen, and the Harness team didn't get code execution in their lab. With the default heap the write was silent across thousands of triggers; only after shrinking TLABs (-XX:TLABSize=2048 -XX:-ResizeTLAB) did they get a SIGSEGV in G1. Calif's own exploit is spray-and-retry heap grooming, and they say it is sensitive to RTT and jitter. On Lobsters someone called it "a nothingburger" next to Log4Shell, and the EPSS score quoted there was 0.00734.
I still think it's the worst JVM bug since Log4Shell, for two reasons.
First, the preconditions are boring. As a Lobsters commenter pointed out, Traefik, Envoy and HAProxy don't buffer requests by default, and Zen 4 shipped in 2022. And Base64 from strangers arrives before login all the time. Harness lists HTTP Basic headers, JWT segments decoded before the signature check, and SASL exchanges like Kafka SCRAM.
Second, and this is the Log4Shell part, the fix lives in every copy of the runtime. Log4Shell at least came down to one dependency in a pom.xml. Here you have to hunt down every JVM you ship, and most teams have no list of those.
Calif also says finding the bug took about an hour and the first working exploit two more, for under $200 in token costs. They're holding back the per-product chains until people have upgraded. Aikido had seen no confirmed exploitation in the wild as of November 18. I wouldn't bet on that lasting.
Which Java builds are affected
All of these need to be true:
- an affected Java 17, 21, 25 or 27 build (Calif puts the start at 17.0.6)
- x86-64 with AVX-512 VBMI: Intel Xeon Ice Lake, Sapphire Rapids, Emerald Rapids, Granite Rapids; AMD Zen 4 or newer
- the JVM isn't already running with
-XX:UseAVX=2or-XX:-UseBASE64Intrinsics - attacker-controlled Base64 reaching
java.util.Base64.Decoder
ARM64 is not affected. Graviton fleets can relax. EPYC Naples, Rome and Milan are AVX2-only, as one HN commenter noted, so older boxes are out of scope too.
According to Aikido, standard Java 8 and 11 are not affected, with one exception: Oracle lists its Java SE 8 Enterprise Performance Pack build (8u511-perf) as vulnerable.
Fixed builds, from Aikido's roundup:
| Line | Oracle fixed | Temurin fixed | Corretto fixed |
|---|---|---|---|
| 17 | 17.0.21.1 | 17.0.21.1+2 | 17.0.21.1.2 |
| 21 | 21.0.13.1 | 21.0.13.1+2 | 21.0.13.1.2 |
| 25 | 25.0.5.1 | 25.0.5.1+2 | 25.0.5.1.2 |
| 27 | 27.0.1.1 | 27.0.1.1+4 | 27.0.1.1.4 |
EOL lines are affected and don't get their own fix. Move 18-20 to 21, 22-24 to 25, and 26 to 27. Red Hat is expediting RHEL 8, 9 and 10 builds. Zulu and Microsoft number their builds differently, Aikido has no fixed build numbers for Liberica yet, and distro packages carry their own versioning, so check each vendor's advisory and don't eyeball it against Oracle's table.
One Temurin gotcha: Adoptium publishes the binaries per platform as testing finishes, and the Docker Hub images only follow once every binary for that version is out. Your FROM eclipse-temurin:17-jre rebuild may pull the old JRE for a while.
Check 1: does the CPU even have the instructions
This one is from Calif's writeup:
grep -m1 '^flags' /proc/cpuinfo | tr ' ' '\n' | grep -E '^(avx512f|avx512bw|avx512vbmi|bmi2)$' All four flags need to show up. If any are missing, this host can't hit the vulnerable path. But your autoscaling group can land the same image on a Sapphire Rapids node tomorrow, so don't treat a clean dev box as proof about production.
Check 2: every runtime on disk, including jlink'd ones
Every JDK, JRE and jlink-built runtime has a release file at its root. Reading those files covers bundled runtimes without executing anything:
sudo find / \( -path /proc -o -path /sys \) -prune -o -type f -name release -print0 2>/dev/null \ | xargs -0 grep -H -E '^(IMPLEMENTOR|JAVA_VERSION|JAVA_RUNTIME_VERSION)=' 2>/dev/null This is where the surprises turn up: /opt/<vendor>/jre, jpackage'd desktop tools, an IDE's bundled runtime, the copy your monitoring agent brought along.
Check 3: what's actually running
The binary on disk doesn't matter if the process still has the old one loaded. Harness's advice is to check the running process rather than trust the image tag. Finding processes by name misses custom launchers, so match on libjvm.so instead:
for m in $(sudo grep -l libjvm.so /proc/[0-9]*/maps 2>/dev/null); do pid=$(echo "$m" | cut -d/ -f3) echo "$pid $(sudo readlink /proc/$pid/exe) $(sudo grep -m1 -o '/[^ ]*libjvm.so' /proc/$pid/maps)" done For each hit, jcmd <pid> VM.version tells you the build, and jcmd <pid> VM.flags -all | grep -E 'UseBASE64Intrinsics|UseAVX' tells you whether the fast path is on. To see what a particular binary defaults to on this CPU:
/path/to/bin/java -XX:+PrintFlagsFinal -version | grep -E 'UseBASE64Intrinsics|UseAVX' UseBASE64Intrinsics true and UseAVX at 3 on a VBMI-capable host means you're exposed until you patch.
Check 4: container images, distroless included
Distroless images don't have a shell or find, so don't bother with docker exec. Export the filesystem and read the release file out of the tarball:
for img in $(docker image ls --format '{{.Repository}}:{{.Tag}}' | grep -v '<none>'); do cid=$(docker create "$img" true 2>/dev/null) || continue # dummy cmd, never run; covers images with no CMD docker export "$cid" | tar -xOf - --wildcards --no-anchored 'release' 2>/dev/null \ | grep -E '^(IMPLEMENTOR|JAVA_RUNTIME_VERSION)=' | sed "s|^|$img: |" docker rm "$cid" >/dev/null done For Kubernetes, kubectl get pods -A -o jsonpath='{..image}' | tr ' ' '\n' | sort -u gives you the image list to feed into the same loop.
Harness also published a runtime checker script that compares local JDKs against the advisory builds and checks the CPU and flags; exit code 2 means a JDK is still exposed. Their post shows it as curl | bash. Read it before you pipe it into a root shell on prod.
Stopgaps if you can't patch tonight
Calif lists both of these flags as workarounds and still tells everyone to update right away.
-
-XX:-UseBASE64Intrinsicsforces the pure-Java decoder. This is the targeted fix. Harness notes it costs performance on Base64-heavy workloads. -
-XX:UseAVX=2turns off AVX-512 for the whole process. Red Hat recommends it, but it changes vectorization everywhere. An ARM64 JVM won't start with this flag, so don't bake it into a multi-arch image'sJAVA_TOOL_OPTIONS.
Strict Base64 validation in front of the decoder (length a multiple of 4, alphabet only, padding only at the end) is a backstop. Harness notes that a loose character-class check doesn't catch the inputs that overflow.
Switching to ZGC isn't a fix either. The only exploit discussed so far targets G1, but as one Lobsters commenter pointed out, the bug isn't in G1 and can still be used for denial of service under ZGC.
My order would be: internet-facing services that decode Base64 before auth (SSO, CI, artifact repos) first, then image rebuilds, then the appliances you'll have to open vendor tickets for.
Did your scan find a Java runtime you didn't know you were running? Where was it hiding? And if you think the Log4Shell comparison is overblown, make your case in the comments.
Sources
- Calif, Java4Shell: Write once, RCE everywhere
- Aikido, Java4Shell: pre-auth RCE flaw in Java 17 puts servers at risk
- Harness, Past the end of the buffer: CVE-2026-114938
- OpenJDK advisory: https://openjdk.org/groups/vulnerability/advisories/2026-11-17
- Oracle CPU November 2026: https://www.oracle.com/security-alerts/cspunov2026.html
- Discussion: Lobsters, Hacker News



