One equals sign, 62 bytes past the heap, and a score nobody agrees on

In short
Key takeaways
- OpenJDK advisory 2026/11/17 lists CVE-2026-114938 in hotspot and compiler at CVSS 3.1 8.1, affecting versions 17, 21, 25 and 27.
- A 65-byte input of 64 A characters plus = makes the AVX-512 decoder write 64 bytes into a 50-byte buffer, putting 62 bytes past its end.
- JDK-8393212, committed under embargo on 14 October 2026, changes four lines and is in OpenJDK 27.0.1.1 but not in the public 17u, 21u and 25u update repos as of 17 November.
- Calif claims pre-auth RCE in Keycloak, Jira, Confluence, JFrog Artifactory, Jenkins and PingFederate, and post-auth RCE in Kafka, Solr, Elasticsearch, Gerrit and Spinnaker; other researchers reproduced crashes and memory leaks, not code execution.
- The New York Times names the model GPT-6 Cyber while Calif's own writeup names Codex; both say the discovery was AI-assisted.
- CISA said it did not link the flaw to any recent outages, and the CVE record lists it as not in the Known Exploited Vulnerabilities catalogue.
A Base64 string of 64 A characters followed by one equals sign writes 62 bytes past the end of a heap array on an affected OpenJDK build. Java rejects the string with IllegalArgumentException after the write has happened. The bug is CVE-2026-114938, disclosed publicly on 17 November 2026 under the name Java4Shell. Oracle and OpenJDK patched it that day. Calif, the firm that found it, says it escalates to remote code execution in eleven widely deployed products. Independent researchers have so far reproduced crashes and memory leaks, not code execution.
That gap between a verified memory bug and a claimed chain to code execution is the most useful thing about this disclosure. There is a CVE record, an OpenJDK advisory, a four-line fix in public source, a CISA alert, and a proof of concept the researcher published. The escalation claims sit apart from all of that, and they come from one source.
The tail that underflows
OpenJDK HotSpot's C2 JIT compiles hot methods to native code. On x86-64 processors with AVX-512 VBMI, Base64 decoding runs through a hand-written assembly intrinsic rather than the generic Java loop. The intrinsic handles 256-byte blocks, then 64-byte blocks, then whatever remains in a tail under 64 bytes. For that last piece it uses evmovdqub, the intrinsic form of VMOVDQU8, which takes a 64-bit mask deciding which bytes get read and written.
The mask is computed from the input and output lengths, then adjusted when the input ends in padding. Calif's writeup sets out the logic in pseudocode. The input mask covers the input length. The output length is three quarters of the input length, rounded down, and drops by one for each trailing equals sign. The output mask is derived from that result with a bit operation that keeps the low bits.
The failure comes when padding makes the output length negative. For tails shaped like =, ==, X=, X==, XY= or ===, the output length underflows to minus one or minus two. The bit operation that builds the output mask treats its count as unsigned, so a negative value leaves every bit set. The mask becomes all ones and evmovdqub writes the full 64 bytes instead of the short tail.
Run the numbers on the published proof of concept. A 65-byte input, 64 A characters plus one equals sign, makes the runtime allocate a 50-byte buffer. The intrinsic writes 64 bytes starting at offset 48. That ends at byte 112, which is 62 bytes beyond a 50-byte buffer. The proof of concept reports REPRODUCED when 62 bytes at offsets 50 through 111 have changed. Java then throws its exception, but the corruption has already happened.
On older builds the trigger is longer: 128 A characters plus an equals sign. The arithmetic underneath is the same.
Where the bug exists and where it does not
Scope is well established. The faulty decoder entered OpenJDK 17.0.6 in January 2023, backported from JDK 18 under JDK-8296735. The advisory of 17 November 2026 lists versions 17, 21, 25 and 27. The bug needs x86-64 with AVX-512 VBMI, which Intel shipped in Ice Lake, Sapphire Rapids, Emerald Rapids and Granite Rapids, and AMD in Zen 4 and later.
Endor Labs reproduced the 62-byte overflow under an AVX-512 emulator on Temurin 17.0.10 and 17.0.12 and cross-checked the scope against Calif's writeup. Upwind reported the same 62-byte case on real hardware.
Several environments are out of scope. Java 8 and 11 do not carry the backport. ARM64 machines, including Amazon Graviton and Apple silicon, do not have the instruction. Older x86 chips before Ice Lake or Zen 4 lack AVX-512 VBMI. The MIME decoder and Base64 libraries outside java.util.Base64 use different code. Advice to patch this specific bug on a Graviton fleet would be wrong.
A four-line fix and a patch gap
JDK-8393212 was committed under embargo on 14 October 2026 under the bland title Base64 optimization improvements. The CVE number was reserved on 22 October. The fix changes four lines. Endor Labs identified it by diffing the tags jdk-27.0.1-ga and jdk-27.0.1.1-ga, which puts the location beyond doubt.
OpenJDK published its advisory on 17 November 2026, last updated at 17:44 UTC. Oracle released its November Critical Security Patch Update at 1:00 pm Pacific time the same day. CISA published an alert titled Critical Java Vulnerability Requires Immediate Patching.
There is a catch in the word immediate. OpenJDK 27.0.1.1 carries the fix, but as of 17 November the public jdk17u, jdk21u and jdk25u update repositories did not, even though those are the versions most deployments run. Vendors ship on their own schedules. A Hacker News comment offered a stopgap of disabling Base64 intrinsics through a JVM flag, which removes the faulty code path at a performance cost. That is a workaround from a forum, not a vendor instruction, and it needs testing before use.
Two scores, no agreement
The authoritative record and the discoverer disagree, and the disagreement is worth keeping rather than smoothing over.
| Score | Scoring version | Assigned by | Attack complexity |
|---|---|---|---|
| 8.1 High | CVSS 3.1 | Oracle, as CNA, in the CVE record | High; the description says difficult to exploit |
| 9.2 Critical | CVSS 4.0 | Calif, with SentinelOne and Aikido | Lower, under the discoverer's threat model |
The CVE record gives CVSS 3.1 base score 8.1, High, with vector AV:N/AC:H/PR:N/UI:N/S:U/C:H/I:H/A:H. It was published on 17 November 2026 at 21:32:03 UTC, reserved on 22 October, and classified as CWE-122, heap-based buffer overflow. Its description calls the flaw difficult to exploit and says an unauthenticated attacker with network access through multiple protocols could compromise the system. CISA's companion SSVC assessment records exploitation as none, automatable as no, and technical impact as total.
Calif, joined by SentinelOne and Aikido, reports CVSS 4.0 at 9.2, Critical. The headline number comes from the discoverer, not from the CNA record.
Both numbers can be true under different scoring versions and different assumptions about how hard the attack is. That does not make them the same claim. The 8.1 assumes high attack complexity; the 9.2 assumes a working chain that lowers it. Until those chains are public, the 9.2 is an assertion about a capability nobody outside the discoverer has demonstrated. This post carries both scores and does not reconcile them.
The products, and what has been shown
Calif's escalation claim names specific software. Pre-authentication RCE is claimed in default configurations of Keycloak, Jira, Confluence, JFrog Artifactory, Jenkins and PingFederate. Post-authentication RCE is claimed in Kafka, Solr, Elasticsearch, Gerrit and Spinnaker. The firm is holding the weaponised chains private until users have patched.
| Claim | Source | Status |
|---|---|---|
| 62-byte heap overflow in the AVX-512 decoder | Calif disclosure, fix diff, Endor Labs, Upwind | Verified, independently reproduced |
| CVE-2026-114938 is this bug | OpenJDK advisory, CVE record | Confirmed |
| Pre-auth RCE in six named products | Calif, relayed by Aikido, Endor Labs and CISA | Reported, chains withheld, not reproduced |
| Found by an AI in about an hour for under 200 dollars | Calif, with the New York Times naming the model | Self-reported, model name contested |
| CVSS 9.2 versus 8.1 | Calif on one side, CVE record and Oracle on the other | Both primary, unresolved dispute |
The technical route described in the writeup runs through several steps. The fixed-value overwrite becomes a bounded chosen-value write by corrupting a buffer length, a technique borrowed from JavaScript engine exploits. That write becomes a heap reader, which Calif calls JavaBleed, and from there the chain reaches the Code Cache and JVM hooks. Remote heap grooming uses TCP acknowledgement timing to control allocation order and prefers a low, stable round-trip time. The writeup says the researcher rented a VM in the same region as the target to get that, describing it as using money to solve the problem. A proxy that buffers whole request bodies, as nginx does by default, disrupts some grooming profiles but not all. Elasticsearch runs from one complete request and WebSocket frames bypass the buffering.
None of that chain has been independently reproduced. Aikido notes that other researchers reproduced crashes and memory leaks rather than code execution. Endor Labs verified the memory corruption and the fix, not the eleven product chains. CISA's warning says the flaw could allow unauthenticated remote code execution without naming products. The defensible statement is that the overflow is real and the escalation is claimed.
An AI found it, and the model name is contested
Calif's writeup describes the discovery. Researcher Quang Luong told Codex to audit JIT compilation, intrinsics and cryptography. After about 20 minutes of prompting and discarding poor results, the agent found the bug in roughly an hour. A first working exploit took two more hours. Token cost was under 200 dollars.
The New York Times, in a story by Dustin Volz and Dylan Freedman on 17 November 2026, reports that Calif used OpenAI's recently released GPT-6 Cyber model. It quotes senior researcher Bruce Dang saying Java is the beating heart of the modern economy. Calif's own account names Codex. Both sources agree the discovery was AI-assisted. The specific model is contested, and this post does not pick one. The one-hour discovery time and the 200-dollar cost are Calif's figures, with no independent audit.
The New York Times also reports that Dang and colleagues coordinated with intelligence agencies in the Five Eyes alliance and with vendors including Google and Microsoft. That comes from one outlet and a named source.
Endor Labs says Calif published after the flaw leaked ahead of its planned disclosure. Calif's page does not mention a leak. That detail rests on a single secondary source.
The outage story does not survive checking
A viral version of this event ties Java4Shell to a cluster of unrelated outages, including an FAA ground stop, Microsoft 365, the NSA, Apple and JPMorgan. The record does not support that.
The NSA's internal network went down on 17 November 2026, the same day as the disclosure. NSA described it as temporary degradation during scheduled routine maintenance. The New York Times said it was not clear whether the disruption related to the Java bug. Reuters reported on 18 November that CISA did not link the flaw to any recent outages. Rubio, speaking as secretary of state and national security adviser, called the incidents unsettling but unrelated.
Governments did warn. Reuters names the United States, South Korea, Germany, the United Kingdom, China's Tencent and a Dutch agency, and reports an emergency meeting at South Korea's presidential office. It cites a 2021 JetBrains survey putting Java at 53 percent of programming language use in South Korea against 30 percent globally, and a Korean government survey putting it at 72 percent in the public sector.
On exploitation in the wild, the sources agree there was none as of 17 and 18 November. The CVE record lists the flaw as not in the Known Exploited Vulnerabilities catalogue, CISA's SSVC assessment records exploitation as none, German and Dutch agencies reported no known successful attacks, Aikido reported no confirmed cases on 18 November, and Endor Labs reported none on 17 November.
Where this could be wrong
The escalation claims are the largest uncertainty. The chains for the eleven named products are withheld. Independent researchers reproduced crashes and memory leaks, not code execution. If those chains are eventually published and reproduced, the product claims move from reported to verified. If they are not, the 9.2 score loses the assumption that supports it.
The AI details are self-reported. The one-hour discovery, the extra two hours for a working exploit, and the 200-dollar token cost come from Calif. The model name differs between the New York Times and Calif's writeup, and this post does not resolve it.
The New York Times article was partly paywalled; only its opening paragraphs were used. Oracle's patch page and the New York Times blocked direct retrieval, so those facts came from the CVE record, the CISA alert, the OpenJDK advisory, Reuters and vendor blogs. The proof of concept was read but not run locally. The 62-byte figure is corroborated by Endor Labs and Upwind, so that finding does not rest on one account.
The claim that the flaw leaked before planned disclosure comes from one secondary source and is absent from Calif's page. The section above argues against a causal link between the disclosure and the same-day outages; that argument rests on explicit statements from CISA, Reuters and the New York Times, not on an absence of evidence alone.
This post was converted from a research dossier without fresh reporting. It is shorter than the house target because the underlying evidence is thin in exactly the places that matter most: the exploitation chains and the AI cost figures. That limit is stated here rather than filled with padding.
On this page
Sources
- Java4Shell: Write once, RCE everywhereCalif Research, 2026
- OpenJDK Vulnerability Advisory: 2026/11/17OpenJDK, 2026
- CVE-2026-114938 recordCVE Program, 2026
- Critical Java Vulnerability Requires Immediate PatchingCISA, 2026
- A.I. Helps Find Major Security Bug in JavaThe New York Times, 2026
- Governments warn Java flaw leaves computer systems open to hackersReuters, 2026
- Java4Shell: Pre-Auth Heap Overflow in OpenJDKEndor Labs, 2026
- Java4Shell RCE: Affected Versions and FixAikido, 2026
- Defending Against Java4ShellSentinelOne, 2026
- RHSB-2026-018Red Hat, 2026
Frequently asked questions
What is Java4Shell CVE-2026-114938?
Java4Shell is the name given to CVE-2026-114938, a heap-based buffer overflow in the hand-written AVX-512 Base64 decoder in OpenJDK HotSpot. A malformed Base64 string makes the decoder write 62 bytes past the end of a heap array. OpenJDK and Oracle patched it on 17 November 2026. The overflow crashes the JVM on its own; escalation to code execution is a separate, unreproduced claim.
Which Java versions are affected by Java4Shell?
The OpenJDK advisory of 17 November 2026 lists versions 17, 21, 25 and 27. The faulty code entered OpenJDK 17.0.6 in January 2023 as a backport from JDK 18. It runs only on x86-64 processors with AVX-512 VBMI. Java 8 and 11 are out of scope, as are ARM64 systems such as Amazon Graviton and Apple silicon, older x86 chips, the MIME decoder and Base64 libraries outside java.util.Base64.
Is Java4Shell remotely exploitable for code execution?
Calif claims pre-authentication RCE in default configurations of Keycloak, Jira, Confluence, JFrog Artifactory, Jenkins and PingFederate, and post-authentication RCE in Kafka, Solr, Elasticsearch, Gerrit and Spinnaker. The weaponised chains are withheld. Independent researchers reproduced crashes and memory leaks, not code execution. CISA said the flaw could allow unauthenticated remote code execution without naming products, so the escalation remains a claim rather than a verified result.
Why does Calif score Java4Shell 9.2 while Oracle scores it 8.1?
The CVE record, with Oracle as the CNA, assigns CVSS 3.1 base 8.1 High and describes the flaw as difficult to exploit because attack complexity is high. Calif reports CVSS 4.0 9.2 Critical. The two figures use different scoring versions and different assumptions about how hard the attack is. Both are on the record; neither cancels the other, and the higher number rests on exploit chains that have not been published.
Which AI model found the Java4Shell bug?
Sources disagree on the name. Calif's own writeup says researcher Quang Luong directed Codex to audit JIT compilation, intrinsics and cryptography, finding the bug in about an hour for under 200 dollars of tokens. The New York Times on 17 November 2026 reports the model as OpenAI GPT-6 Cyber. Both agree the discovery was AI-assisted. The one-hour time and the cost are self-reported and have not been independently audited.
Was Java4Shell linked to the November 2026 outages?
No causal link is supported by the record. The NSA internal network went down on 17 November 2026, the day of disclosure, and NSA called it temporary degradation during scheduled routine maintenance. The New York Times said it was unclear whether that disruption related to the bug. Reuters reported on 18 November that CISA did not link the flaw to any recent outages, and officials described the incidents as unrelated.