Java4Shell: Oracle scores CVE-2026-114938 8.1, Talos and the finder 9.2; finder says 1 hour to find, under $200
Oracle scores CVE-2026-114938, a flaw in OpenJDK's x86-64 Base64 decoder, 8.1 under CVSS 3.1; Cisco Talos and the finder, Calif.io, score it 9.2 under CVSS 4.0. No saved page says it is exploited, and the finder says it took 1 hour to find and two more to a first exploit for under $200 of tokens.
By Parminder Kumar Sharma · LinkedIn · · 17 min read
One flaw, three pages and two scores: 8.1 and 9.2 are different versions of CVSS
Three saved primary pages describe CVE-2026-114938, a memory corruption flaw in the x86-64 Base64 decoder of OpenJDK's HotSpot virtual machine, which its finder, Calif.io, calls Java4Shell. Oracle's advisory and CVE record score it 8.1 under CVSS 3.1, with attack complexity High. The Cisco Talos advisory and the finder's write-up score it 9.2 under CVSS 4.0. The 1.1 point gap (derived: 9.2 less 8.1 = 1.1) sits between two versions of the scoring system, so it is not a measure of how much the authors disagree about the flaw. No saved page says the flaw is being exploited.
The finder also publishes a pace. Its write-up says one of its researchers needed "an hour to discover the vulnerability, and two more hours to develop the first working exploit" and that "The total token cost was less than $200." That is 3 hours (derived: 1 + 2) and at most about $66.67 of tokens an hour (derived: $200 / 3 h, an upper bound because the page says "less than"). It is the finder's own claim about its own work: no saved page measures it, and the cost is tokens only.
What that does not establish. It does not establish that anyone has exploited the flaw, that every Java service is reachable (the pages name a chip condition and the finder says there is no universal exploit), how many systems are exposed, or which fixed build replaces which affected one in each distribution: no saved page prints a fixed version number. This site reports what the pages print, works out the figures it can, and labels every figure it worked out "derived".
Whose account this is, and what each seller gains. Calif.io is a security research and consulting firm; its page says its researchers used AI to find and weaponise the bug and frames the work as the "centaur hacker", and a finding like this also advertises that approach. The write-up names the AI tool as Codex, which OpenAI makes. Oracle sells Java support and assigned and scored its own CVE. Cisco Talos writes the advisory and lists Snort rule coverage for subscribers, so its advisory also describes products Cisco sells. The finder's earlier PHP2Shell disclosure, which this site covered, was a separate flaw in PHP.
What the pages say the flaw is
The finder's write-up says OpenJDK uses specialised x86-64 instructions, AVX-512 VBMI, to decode Base64 faster, and that the bug "sits in the hand-rolled assembly code that handles the final few characters". In plain words: when the JVM has compiled the decoder into that fast routine and a malformed input ends in padding, the routine miscalculates how many bytes to store. The write-up says the output length is computed as a negative number and "becomes a full 64-byte write", so the store runs past the end of the result array. It adds that Java then rejects the input, "but by then neighboring objects may already be corrupted".
Talos describes the trigger as "sending a malformed Base64 input to affected JVMs". The finder gives the length of the smallest input in words, so this site does the same and prints no input: 64 characters followed by one padding character, and 128 characters plus the padding character in older OpenJDK 17, because the finder says the length threshold to reach the code is higher there. The finder says ordinary features such as authentication and file uploads can expose a Base64 decoder, and that a Java program opening a TLS connection to a malicious server can be compromised even if it offers no service, a client side case a second researcher shared with it.
The finder's byte counts, as printed and as this site re-derived them (derived column)
| Figure | The write-up prints | Derived by this site |
|---|---|---|
| Result array for the 65 character input | 50 bytes | ceil(65 / 4) = 17; 17 x 3 = 51; 51 less 1 padding = 50 |
| Where the 64 byte store starts | offset 48 | 48 is inside the array; bytes 48 and 49 are in bounds |
| Bytes written past the end | 62 bytes | 48 + 64 = 112; 112 less 50 = 62 |
What the finder says can follow is a heap leak "in the style of Heartbleed" and remote code execution; Talos says the flaw "may lead to unauthenticated remote code execution". The write-up says the bytes written are fixed values, that exploitation therefore looked difficult, and that it overcame this with a general technique and per product work. This site does not describe that technique. Both pages call the result critical; neither page prints a CVSS 4.0 vector.
Three records of one flaw, side by side
Each page answers a different question. Oracle's page is a patch advisory with a risk matrix: it names the Oracle builds it calls affected and scores them, and its row for this CVE does not mention a chip condition. Talos's page is a threat advisory: it names the chip condition, says none was observed in the wild "yet" and lists Snort coverage. The finder's page is the discoverer's own story and technical write-up: it names the products, the distributions and the AI claim. Where a page is silent the diagram says so rather than filling the cell.
What the pages state about the questions a reader asks, and what none of them states
| Question | Stated, and by whom | Not stated |
|---|---|---|
| Is it exploited? | Talos: none observed in the wild yet, probing expected. CISA-ADP entry in the CVE record: exploitation none, at 21:53:20 UTC on 17 November | Anything after those times; any report by anyone else |
| Which builds are affected? | Oracle: eight build numbers in its row across Java SE and two GraalVM products (derived count). Talos and the finder: OpenJDK 17.0.6 and later. The finder: ten distributions are vulnerable because they are built from OpenJDK | A fixed build number for any distribution; why Oracle lists a Java 8 build |
| Which hardware? | Talos and the finder: x86-64 with AVX-512 VBMI. The finder: ARM64 is not affected, and lists Intel and AMD families | Oracle's page and the CVE record print no chip condition |
| Which products? | The finder: 11 named products with RCE "in their default configs", non-exhaustive, and sample Struts and Spring Boot applications | Any vendor's statement about its own product; Cisco says it will provide updates shortly |
| How hard is it? | Oracle: "Difficult to exploit", attack complexity High. The finder: no universal exploit, each target needs its own steps | A measure of how many targets are reachable by the general technique |
| How many hosts? | Talos: a large potential attack surface, on premise and cloud | A count of exposed systems |
The order of events, to scale in UTC
Talos's page prints "Tuesday, November 17, 2026 20:53" with no time zone, which on its face puts it before the CVE record's 21:32:03 UTC publication. The page's own data says otherwise: its published time is 01:53:03 UTC on 18 November, and 20:53 plus 5 hours is 01:53, so the printed time reads as UTC-5 (derived). On that reading the Talos advisory came 4 h 21 min after the CVE record (derived: 01:53:03 less 21:32:03 = 4 h 21 min 0 s), not before it.
The finder's page says OpenJDK released an advisory and Oracle a patch "at 1:00pm PT on Nov 17, 2026". Pacific time on that date is UTC-8, because US clocks went back on Sunday 1 November 2026 (derived: that day is a Sunday), so 1:00pm is 21:00 UTC. That is an assumption about which clock the finder meant, so the diagram draws it as a hollow dot. On it, the CVE record followed the patch by 32 min 3 s (derived: 21:32:03 less 21:00:00). The record itself carries a CISA entry timestamped 21:53:20 UTC, 21 min 17 s after publication (derived), with exploitation none, automatable no and technical impact total, and it was updated again at 04:49:49 UTC on 18 November. The id was reserved at 04:13:03 UTC on 22 October 2026, 26 days 17 h 19 min before publication (derived).
A nickname and two scores do the work of an assessment
"Java4Shell" is the finder's name, and Talos prints it in quotation marks after the word "dubbed". The name invites a comparison with other famous Java flaws that none of the saved pages draws, and the CVE record, Oracle's advisory and the CVE id carry no nickname. A score does the same job from the other side: "CRITICAL" in the finder's page and 8.1 HIGH in the CVE record are labels, and the CVE record's own wording for the same flaw is "Difficult to exploit". The finder's wording is different again: it says there is no universal exploit and that different steps may be required depending on how Base64 is used.
None of those labels answers the questions a defender can check: is the JVM on a chip with AVX-512 VBMI, does the service pass attacker controlled Base64 to the decoder, has the build been updated, and would anyone notice an attempt. The CISA entry in the CVE record adds two machine readable answers, attributed to CISA-ADP and its coordinator role: exploitation none and automatable no at 21:53:20 UTC. Neither is a promise about later, and "automatable no" is not the same as "cannot be done by hand".
Where the sources differ
Each difference is reported and none is resolved; this site does not guess which page is right.
- Score. Oracle: CVSS 3.1 base 8.1, vector as the CVE record prints it (see the technical section). Talos and the finder: CVSS 4.0 9.2, no vector printed. Two versions of the system, so the numbers are not directly comparable.
- Versions. Oracle's row lists its own build numbers, including a Java 8 build, 8u511b33-perf, and the finder repeats that list "according to Oracle". Talos and the finder say OpenJDK 17.0.6 and later. No saved page explains how a Java 8 build fits a flaw that starts at 17.0.6, and none prints which fixed build replaces an affected one.
- Chip condition. Talos and the finder: x86-64 with AVX-512 VBMI. Oracle's row and the CVE record print no chip condition at all.
- How hard. Oracle: difficult to exploit. The finder: pre-auth RCE in six products and post-auth RCE in five, in default configurations, and a client side case, but no universal exploit. Both are reported as printed.
- Exploitation. Talos: none observed yet, probing expected. The finder: it holds back its product chains "until everyone has had a chance to upgrade" and says it will publish them later, and it publishes a reproducer that checks for the out of bounds write. This site prints neither. No page says anyone else has exploited the flaw.
- Mitigation flags. The finder prints two flags: one disables the hardware accelerated Base64 intrinsics and one disables AVX-512. Talos prints one, with a leading equals sign where a hyphen is expected (see the technical section). Both pages warn that the AVX flag stops OpenJDK starting on ARM.
- Credits. Oracle's credit line for this CVE names two people; the finder's page credits others for research, exploit work and artwork. This site prints no names.
- AI. Only the finder's page mentions AI. Oracle's row, the CVE record and the Talos advisory body do not.
The finder's account of how it was found, and what it leaves out
The finder's write-up says its researcher "told Codex to look at" the JIT compiler, intrinsics, widely reachable APIs and cryptography, that most early results were slop, and that "around 20 minutes" of prompting and ignoring slop produced a finding in the Base64 decoder. The overview says AI also helped to weaponise the flaw: Calif spent "initial effort building pipelines and strategies", then directed agents to use them, and "most of the later chains are built almost autonomously with AI". It also says its researchers keep the human part: the page's own phrase is the "centaur hacker, part human, part AI".
What the cost claim covers. "Less than $200" is a token cost for the work described, which the page puts at 3 hours (derived). It is not stated whether the pipelines and strategies built before the work are counted, what the researcher's time costs, or how many attempts failed. The write-up also says its discovery came after the researcher had audited other code for an earlier project, so the one hour is a claim about the last stage. None of this is independently measured on any saved page.
Products, as the finder's claim. The write-up lists pre-auth RCE in Keycloak, Atlassian Jira, Atlassian Confluence, JFrog Artifactory, Jenkins and PingFederate, and post-auth RCE in Apache Kafka broker, Apache Solr, Elasticsearch, Gerrit and Spinnaker, that is 11 products (derived count), "in their default configs", and calls the list non-exhaustive. No vendor statement is saved, so this site cannot say whether a customer's configuration is exploitable. The write-up's Limitations section says a buffering reverse proxy affects some of its chains but "Request buffering is not a general mitigation" and that one listed product needs only a single complete request: this site does not treat buffering as a control.
A reader should also keep the incentives in view: a finding that took three hours and under $200 is a strong message for a firm that sells AI assisted security work, and it is also the kind of message that should be checked rather than repeated. Oracle, Talos and the CVE record say nothing about how the bug was found.
What the pages print, copied as printed
This section copies the defensive items the saved pages print, each attributed in its caption, and nothing else. The finder's write-up also prints a bounded trigger program and a pseudocode listing; this site does not copy them, because a copy would be a reproducer. The pages print no file hashes, network indicators or log lines.
Oracle's risk matrix row. The row below is this site's layout of the table cells; the values are Oracle's. Note 1 is printed under the table and says the flaw is reached by supplying data to APIs, such as through a web service, and not through untrusted Java Web Start applications or applets. A defender reads it as: server side Java that decodes data from requests is in scope.
CVE ID: CVE-2026-114938
Product: Oracle Java SE, Oracle GraalVM for JDK, Oracle GraalVM Enterprise Edition
Component: Compiler
Protocol: Multiple
Remote Exploit without Auth.?: Yes
CVSS VERSION 3.1: Base Score 8.1; Attack Vector Network; Attack Complex High; Privs Req'd None; User Interact None; Scope Unchanged; Confidentiality High; Integrity High; Availability High
Supported Versions Affected:
Oracle Java SE: 8u511b33-perf, 17.0.21.0.3, 21.0.13.0.3, 25.0.5.0.3, 27.0.1.0.3
Oracle GraalVM for JDK: 17.0.21.0.3, 21.0.13.0.3
Oracle GraalVM Enterprise Edition: 21.3.20.0.3
Notes: See Note 1
This vulnerability can only be exploited by supplying data to APIs in the specified Component without using Untrusted Java Web Start applications or Untrusted Java applets, such as through a web service.
The CVE record. The vector is the one the CVE record carries for Oracle's score. The CWE and the SSVC entry come from the CISA-ADP container in the same record, not from Oracle.
vector: CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:H/I:H/A:H
baseScore: 8.1 (HIGH)
CWE (CISA-ADP container): CWE-122 Heap-based Buffer Overflow
SSVC (CISA-ADP container, role CISA Coordinator, timestamp 2026-11-17T21:53:20.487730Z):
Exploitation: none
Automatable: no
Technical Impact: total
datePublished: 2026-11-17T21:32:03.183Z
dateUpdated: 2026-11-18T04:49:49.253Z
The chip condition. The finder says the vulnerable code is hardware specific, lists the families below, and says the command lists the CPU flags on Linux and that "The output must include all four flags". It reads one line of a system file and changes nothing. Talos names the condition as AVX-512 VBMI support enabled.
Intel Xeon: Ice Lake, Sapphire Rapids, Emerald Rapids, Granite Rapids
AMD EPYC and Ryzen: Zen 4 or newer
grep -m1 '^flags' /proc/cpuinfo | tr ' ' '\n' | grep -E '^(avx512f|avx512bw|avx512vbmi|bmi2)$'
The two mitigation flags and Talos's variant. The finder says to start OpenJDK with either flag; the first disables hardware accelerated Base64 and the second disables AVX-512, to be applied only on x86-64 hosts. Talos prints its flag with a leading equals sign and two X characters where the finder's reads hyphen, X, X: it looks like a typo, and this site has not tested either. Both pages warn that ARM hosts will not recognise the AVX flag and OpenJDK will fail to start there. Neither page says what the flags cost in speed.
The finder prints: -XX:-UseBASE64Intrinsics
The finder prints: -XX:UseAVX=2
Talos prints: =XX:UseAVX=2
Detection. Talos says its rules detect malformed Base64 data in specific HTTP request fields, that the flaw is implementation specific, and that the community ruleset is live. The ids it prints are below. A rule that watches HTTP fields would not see the client side case through a TLS connection.
Snort2: 1:67502 - 1:67513
Snort3: 1:301754, 1:67502 - 1:67513
What to do, in order
Take this with you
This site's suggestions, in the order worth doing them
- Update every JVM you run from your vendor's fixed build, and record the build number next to each host. No saved page prints which build is fixed, so take the number from your vendor's advisory.
- Inventory Java based products on x86-64 hosts: the products the finder lists, anything on Spring Boot or Struts, and anything that opens TLS connections to servers you do not control. Ask each vendor which fixed version replaces which.
- Check hosts for the four CPU flags using the finder's printed command. Hosts without all four, and ARM64 hosts, are outside the condition the pages describe. Record the result per host.
- Where you cannot update yet, apply one of the printed flags on x86-64 hosts only, test it first, and measure speed: the pages say nothing about cost. Do not apply the AVX flag to ARM hosts.
- Put internet facing Java services behind controls that can inspect request fields and keep the logs. Talos lists Snort coverage for malformed Base64 in specific HTTP request fields; it does not cover a client side case.
- Watch for exploitation reports. Talos expects probing soon after disclosure and the finder says its product chains will be published after others have had time to upgrade.
- When someone quotes a score at you, ask which CVSS version it is. 8.1 and 9.2 are the same flaw scored two ways, and neither says it is being exploited.
The question the pages leave open
Every page agrees on the narrow condition: a Java runtime of OpenJDK 17.0.6 or later, an x86-64 chip with AVX-512 VBMI, and attacker controlled Base64 reaching the decoder. None of them says how many of your Java services meet all three, and none lists the fixed build that ends it. Do you know which of your Java services run on a CPU that makes this reachable, and who checked?
Key facts
Sources
- PrimaryJava4Shell: Write once, RCE everywhere, the finder's story and technical write-up: overview, products tested, AI use and cost, affected products and distributions, CPU families and flag check, recommendations. Read in full on 18 November 2026. The page prints no dateCalif Research (Calif.io)accessed 2026-11-18
- PrimaryOracle Critical Security Patch Update Advisory, November 2026: 655 patches, 11 for Java SE; the CVE-2026-114938 row of the Java SE Risk Matrix, Note 1 and the credit statement. Read on 18 November 2026Oracleaccessed 2026-11-18
- PrimaryCVE-2026-114938: Critical vulnerability in OpenJDK JAVA, Threat Advisory printed Tuesday, November 17, 2026 20:53 (page data: 18 November 2026 01:53:03 UTC): CVSS 4.0 9.2, condition, no exploitation observed yet, mitigation, Snort ids. Read on 18 November 2026Cisco Talosaccessed 2026-11-18
- PrimaryCVE-2026-114938 record read from the CVE Services API on 18 November 2026: published 2026-11-17T21:32:03Z, CVSS 3.1 8.1 vector, affected Oracle versions, CISA-ADP CWE-122 and SSVC entry, reference listCVE Program (assigned by Oracle)accessed 2026-11-18


