decrypted · 18 november 2026 · vulnerabilities and patching · uk policy and law · ai and llm security
Java4Shell: a decoding flaw that can let strangers run code on Java servers
Java is the quiet engine under much of what UK organisations run. On 17 November, Oracle's November Critical Patch Update fixed a flaw in it that researchers have named Java4Shell, tracked as CVE-2026-114938 (a CVE being the public reference number every known flaw gets). CISA, America's national cyber security agency, urged immediate patching. The detail matters, because whether you are exposed depends on your processors as much as your software.
What actually goes wrong
Base64 is a way of writing binary data as plain text. Applications use it constantly, for example to carry login details in web requests. Java has a fast path for decoding it, hand-written for processors with a feature called AVX-512 VBMI. Qualys and Calif, who found it, place the bug in how that code treats the final padded group of characters. Picture a sorter told to fill a 64-slot tray when two letters remain: it sweeps the whole tray and spills onto whatever sits beside it. Here a short malformed input writes about 62 bytes past the end of a buffer. Calif says an attacker needs no credentials, and reports demonstrating remote code execution (RCE, running the attacker's own code) on default installs of Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate.
Who is exposed
Three conditions must all hold, according to Rapid7 and Qualys. The runtime must be an affected build: OpenJDK from 17.0.6 onward, including 21, 25 and 27, plus Oracle's Java 8 Enterprise Performance Pack. The processor must be x86-64 with AVX-512 VBMI, which Qualys says means Intel Xeon from Ice Lake onward and AMD Zen 4 onward. ARM64 is not affected. And untrusted data must reach the decoder, such as credentials checked before a login completes. Rapid7 says an affected version alone does not make an application exploitable.
Disputed numbers and exploitation
Oracle rates the flaw 8.1 under CVSS v3.1 (the zero-to-ten severity score), citing high attack complexity. Calif rates it 9.2 under the newer CVSS 4.0. Rapid7 and Qualys list the fixed Oracle builds as 17.0.21.1, 21.0.13.1, 25.0.5.1, 27.0.1.1 and 8u513-perf; other vendors use their own build numbers. As of 17 November, Qualys had seen no public confirmation of exploitation in the wild, and Rapid7 reports none, but Qualys says not to plan around that.
The fix, and the design lesson
Patch, then restart every Java process, because a running JVM keeps the old code. Check bundled runtimes in products and container images. If you cannot patch, both Rapid7 and Qualys give a JVM setting that switches off the fast Base64 decoder at some cost in speed.
The Secure by Design lesson is that the riskiest code is often the cleverest: a performance shortcut in a path every request touches, parsing data from strangers before they have proved who they are. For UK organisations the practical question is inventory. Do you know which products embed Java, and which processors they run on?
Also this week
Policy: the NCSC says cheap tech is now a bill. On 3 November the NCSC, the National Cyber Security Centre, published its annual review. In remarks the NCSC released, chief executive Richard Horne warned of "a real prospect of an unprecedented rise in the number of cyber incidents", not because AI has invented new attacks but because long-tolerated vulnerabilities can now be found and exploited at scale. He added that "we have collectively got away with doing tech on the cheap". The NCSC also says critical national infrastructure should brace for a ransomware attack involving AI-powered malware by 2028. It pointed to £60 million for the first phase of a national Cyber Shield, £70 million for a new vulnerability exposure centre, an AI Cyber Security Centre of Excellence in Manchester and a Cyber Security and Resilience Bill. Money helps, but a flaw that sat in a trusted runtime until a researcher looked is exactly what he means: the fix is dull inventory and patching discipline, not another dashboard.
AI security: a harness built on Claude found 99 flaws in 90 hours. Forescout reported, as covered by Cybersecurity Dive on 12 November, that its AI harness examined ten TCP/IP stacks, the networking code inside IoT (connected everyday devices) and industrial kit, including Contiki, lwIP and wolfIP. It rediscovered 33 known flaws and found 66 new ones, work Forescout estimated would take nine months by hand, for roughly $1,200 to $2,300 in tokens, about 100 times less than a senior vulnerability hunter's salary by its reckoning. About half the new bugs were memory-corruption flaws, the same family as Java4Shell. Of 54 reported to maintainers, 24 are fixed, and the developer of one project, Ethernut, did not respond. The point for buyers is that attackers get the same economics, so ask suppliers now how they will deliver patches for devices that sit in plant rooms for a decade.
Sources
- CISA: Critical Java Vulnerability Requires Immediate Patching
- Rapid7: CVE-2026-114938 reported critical pre-authentication RCE in Java 17
- Qualys: How to find, mitigate and patch Java4Shell (CVE-2026-114938)
- Calif: Java4Shell, write once, RCE everywhere
- NCSC: Cyber chief says UK facing AI-powered attacks on critical infrastructure within two years
- Cybersecurity Dive: AI found widespread IoT vulnerabilities more quickly and cheaply than humans could
If you want help working out where Java hides in your estate, get in touch.
More like this
- The Lords say the government is complacent about cascading crises 16 november 2026
- The second flaw that turned a Check Point foothold into root 14 november 2026
- Barracuda's mail gateway has gone from patched to under attack 11 november 2026
Get the next post by email: subscribe to Decrypted. Double opt-in, unsubscribe any time, or take the RSS feed.
Prefer to listen? Decrypted on Apple Podcasts, or paste the podcast feed into any app.