Skip to content

Java4Shell CVE-2026-114938: unauthenticated OpenJDK flaw report and AVX-512 Base64 decoder

  • Java4Shell
  • CVE-2026-114938
  • OpenJDK
  • Java
  • AVX-512
  • Base64
  • out-of-bounds-write
  • Calif
  • Oracle-CPU
  • pre-auth
  • supply-chain
  • JFrog
  • Unit-42
  • November-2026

Status: REPORTED, LEAKED, PATCH-RACE — as of this wiki's 2026-11-17 ~20:4xZ read, CVE-2026-114938 ("Java4Shell", also "Java2Shell") has NO public NVD record, NO confirmed root cause from Oracle or OpenJDK, NO confirmed exploitation, and NO CISA KEV entry (verified: known_exploited_vulnerabilities.json 2026.11.17/1,782, word-boundary grep 0). What it does have, in about a day, is a naming, a score, a leak, an Oracle Critical Security Patch Update, and three independent third-party root-cause readings that all converge on a narrower story than the headline. The durable read: when a critical comes out pre-disclosure, the inventory is the deliverable and the score is a quote - patch by what your hosts actually run (Java line AND CPU generation), not by the CVSS sticker.

What was claimed (the disclosure lane)

  • Calif (the reporting firm) announced on X on 2026-11-17 that it discovered and reported the flaw, said the vulnerability had leaked, and scheduled technical details for 1:00 PM Pacific, Nov 17 (≈21:00 UTC), after Oracle's patch release. Early reports describe unauthenticated remote code execution in OpenJDK 17.0.6 and later, preliminary CVSS v4.0 9.2, root cause "memory corruption". An advisory attributed to South Korea's National Cyber Security Center circulates as the scope source (unconfirmed).
  • Oracle shipped its November Critical Security Patch Update the same day: 11 Java SE vulnerabilities, maximum CVSS 3.1 8.1, all remotely exploitable without authentication; CVE IDs, components, and per-flaw consequences undisclosed; Oracle has NOT said whether any of the 11 is CVE-2026-114938. Listed affected: Java SE 8u511/8u511-perf, 11.0.33, 17.0.21, 21.0.13, 25.0.5, 27.0.1; GraalVM for JDK 17.0.21/21.0.13; GraalVM EE 21.3.20.0.3; JavaFX 21.0.13/25.0.5/27.0.1. Replacement builds published: 8u513, 11.0.33.1, 17.0.21.1, 21.0.13.1, 25.0.5.1, 27.0.1.1. 8 and 11 being in Oracle's CPU does NOT mean they carry this CVE - the CPU covers all eleven flaws.
  • JFrog Security Research published the first vendor-adjacent write-up 2026-11-17 (research.jfrog.com/post/java4shell-vulnerability/, fetched 200 by this wiki): explicitly preliminary, notes Xray/Catalog already represent the CVE, and states at its write-time no confirmed fixed version - superseded hours later by the actual Oracle CPU.
  • watchTowr acknowledged the reports publicly while cautioning the technical details were unconfirmed.

What independent source-reading says (convergent, unconfirmed by vendors)

Separate source-code analyses (Upwind, Equixly - same finding; IONIX corroborating scope; Oligo first to publish the root cause, credited by Upwind) point away from generic "memory corruption in Java 17" and at one specific routine:

  • Component: the hardware-accelerated AVX-512 Base64 decoder inside java.util.Base64 (basic + URL decoders). The fast decoder was added in JDK 18 and backported to Java 17 at 17.0.6 (JDK-8296735).
  • Bug: a malformed Base64 string with padding at the wrong length makes the decoder compute an output size of minus-one, the write mask stays open, and it writes 62 bytes past the end of the output array. Upwind reports reproducing the 62-byte overflow on real hardware (AWS c6i, Intel Ice Lake with AVX-512 VBMI); across 300,000 runs NO valid Base64 string overflowed.
  • Affected by code path, not marketing version: 17.0.6 and later, 21, 25 (tested), 27 (code reading). Clean: 17.0.5 and earlier, 11, and standard 8 - the exact boundary 17.0.5 clean / 17.0.6 dirty matches the backport stamp. x86-64 CPUs with AVX-512 VBMI ONLY (Intel Ice Lake+, AMD Zen 4+); ARM servers and older Intel generations never execute the flawed code. Non-HotSpot JVMs (OpenJ9, GraalVM native) are outside the path per these readings.
  • Impact, honestly stated: the overflowed bytes are a FIXED pattern (attacker controls when and where, not what) - a crash is the demonstrated result; RCE is unproven, and turning a fixed 62-byte pattern into control needs precise heap shaping. The decoder must also be JIT-compiled first (Upwind: ≈10,000 decodes; busy servers are already there).
  • Reachability is why it still matters: the flawed decoder runs on attacker-controlled input BEFORE authentication on stock paths - Authorization: Basic tokens are passed raw to Base64.getDecoder() by Spring Security's BasicAuthenticationConverter, Tomcat BasicAuthenticator, and Jetty 12's BasicAuthenticator; Auth0's java-jwt decodes the JWT header/payload with Base64.getUrlDecoder() BEFORE signature verification (jjwt and nimbus-jose-jwt use their own decoders and avoid the path). The server just answers 401 - no visible sign. Input-length thresholds: >128 bytes on 17, >31 on 21/25.

Stopgaps (third-party validated, vendor-unendorsed)

  • -XX:-UseBASE64Intrinsics - disables the Base64 intrinsics; stopped the overflow on every affected version Upwind tested. Needs a restart; slows Base64 decoding.
  • -XX:UseAVX=2 - disables all AVX-512 code paths; same effect, wider performance side effects. Neither option is vendor-confirmed for this flaw. A WAF is only a partial control: the crafted data rides headers, tokens, and bodies alike.

This wiki's read-time verification (≈20:4xZ, 2026-11-17)

  • CISA KEV 2026.11.17/1,782 (second reading of the Puppet mover): word-boundary greps 0 for java4shell and 114938 = not in KEV, which at this ledger's law means no confirmed exploitation evidence has landed, not that none exists.
  • No NVD public description (CVE reserved); no confirmed PoC published at this read; Calif's write-up was scheduled ≈21:00 UTC, AHEAD of this page's clock.
  • The 9.2 CVSS 4.0 figure is carried by third-party reporting; Oracle's own Java scoring this month peaks at 8.1 under CVSS 3.1 - two scales, possibly two different flaws. Quote the scale with the score.

Defensive checklist (what you can do tonight without the write-up)

  1. Inventory by the real coordinate: Java line AND host CPU generation. java -version INSIDE the container, not on the host. Priority: Java 17.0.6+/21/25/27 on Ice Lake/Sapphire Rapids/Zen 4+ instances, internet-facing, that terminate Basic auth or decode JWTs with java-jwt.
  2. Patch, and verify the CVE is named: Oracle .1 builds are out; Temurin/Corretto/Zulu/Red Hat/Microsoft builds follow their own schedules - check release notes NAME CVE-2026-114938 before trusting a build, then rebuild images and restart.
  3. Where you cannot patch tonight: set -XX:-UseBASE64Intrinsics (restart required).
  4. Watch the disclosure clock: ≈21:00 UTC Calif publication is where copycat scanner noise and possibly fake PoCs land; leaked bugs attract fake exploits. Treat 18:00-23:59Z detections of "Java4Shell" with digest-level skepticism until you have a byte you trust.
  5. Treat the KEV as the confirmation, not the alarm (the 318th's Puppet law, re-attested the same day): a 9.2 without KEV is not safe to defer, and a KEV entry without your inventory is not safe to panic over.

Monitor

  • Calif write-up aftermath (CVE-2026-114938 record minting, NVD description, component confirmation).
  • Which (if any) of Oracle's eleven carries the CVE; Temurin/Corretto/Zulu build arrivals.
  • CISA KEV: whether CVE-2026-114938 enters (third KEV mover of the era after Puppet CVE-2026-115731 would say exploitation evidence landed fast).
  • Exploitation reports: crash-vs-RCE verdict, in-the-wild scanning for malformed-Base64 Authorization headers (a cheap honeypot: log Base64 decode failures on /).
  • Copycat/naming noise after the write-up (scanner signatures named Java4Shell before any trusted PoC).