On November 17, Java got its own “shell.” CVE-2026-114938, now known as Java4Shell, lets an attacker send one badly padded Base64 value to an affected Java application before any login. The result can be a crash, leaked tokens and keys or, in some conditions, remote code execution.
Surprised? Not really. Shaken? Not at all. This is the third shell in about four weeks, after PHP2Shell and Node2Shell, and Radware AI Xploit Shield customers are already covered:
With almost no effort, they’re deploying five custom signatures for immediate protection before they even patch.
Why “just patch it” is easier said than done
The fix (17.0.21.1, 21.0.13.1, 25.0.5.1, 27.0.1.1 and later) lives in the runtime, so every service on an affected JDK has to be upgraded, retested and restarted. Products that bundle their own Java wait on their vendors, container images need rebuilding, and a day after disclosure several major Java distributions still had no fixed build. Rush it, and you risk an outage of your own making.
Blocking it isn’t simple either. Many security rules block a known attack by matching a pattern taken from the published exploit. Java4Shell doesn’t depend on any particular text. What makes it work is the structure of the input: Base64 padding in the wrong place on a value of the wrong length. Attackers can swap every character in the public example and the attack still works, while a rule built from that example no longer recognizes it.
Want to know more and get a bit nerdier? The full technical details are in Radware’s security advisory on the customer portal (login required): read it here. And for more on how Radware AI Xploit Shield works, start here.
How AI Xploit Shield took it apart
When a vulnerability goes public, most defenses respond with a rule that matches that exact attack. Attackers then change one value, and the rule no longer catches it. AI Xploit Shield looks at how the attack actually works.
- It caught a gap. Shield replayed the attack and found a generic rule that blocked the public proof of concept, but not the same attack with new values. A rule that only blocks the published example isn’t protecting much.
- It targeted the attack’s shape. Shield built protection around what every Java4Shell attempt shares: Base64 padding on a length no legitimate client ever sends. Change the values all you like. The shape gives you away.
- It created five custom signatures. Customers add them to their security policies in the Radware cloud. They can run them in report-only mode first, to make sure nothing legitimate gets flagged, then switch to blocking.
No emergency change window. No 2 a.m. war room.
The patch is the attacker’s roadmap
Attackers now run AI agents that diff every new patch against the old code and turn the diff into a working exploit with no human in the loop. Multi-agent attack frameworks such as ARTEX chain steps that once took skilled teams days. In one government-flagged campaign, a single supply-chain break-in led to sign-in attempts against more than 2,250 downstream customer environments. Zero-day exploits nearly tripled, from about eight a month in 2025 to 22 in August 2026 alone. And exploiting vulnerabilities is now the number one way breaches start, ahead of stolen credentials for the first time in 19 years.
Since the fix shipped, every attacker with an agent has a map to the bug, and a swarm of AI agents can churn out variants faster than any team can write rules for them. AI Xploit Shield’s signatures target the attack’s underlying structure, so new variants run into the same protection.
This is just the beginning
We all know Java4Shell won’t be the last. The next shell is already on its way.
AI Xploit Shield customers aren’t losing sleep though. They’ll upgrade Java on their own schedule, not the attacker’s, patching safely and in peace.