Critical Java Flaw Found by AI Sends Hospitals Racing to Patch
Key Takeaways
- A cybersecurity startup called Calif used OpenAI's unreleased GPT-6 Cyber model to find a critical flaw in Java, the programming language used by about 90% of Fortune 500 companies, as well as banks, hospitals and governments.
- The flaw, nicknamed Java4Shell and tracked as CVE-2026-114938, sits in Java's Base64 decoder. It can crash the Java Virtual Machine, expose sensitive memory, or allow an attacker to run code remotely.
- The flaw had sat unnoticed for 12 years. Calif says it found the bug within about an hour and built a working exploit within two more hours, at a token cost of under $200.
- Oracle, which owns Java, released a critical patch update on 17 November 2026. It has not publicly acknowledged the specific bug, and has not named it in the update.
- The US Cybersecurity and Infrastructure Security Agency (CISA) urged organisations to patch immediately. The American Hospital Association circulated the warning to hospitals on 18 November, and health systems across the US are scanning their networks, medical devices and internet-facing systems.
- Epic, the largest electronic health records vendor, said its systems were not affected. No successful attacks using the flaw have been reported so far, but researchers warn the window to patch is short.
A Twelve-Year-Old Flaw, Found in an Hour
The most startling fact about this story is not that a serious flaw was found in Java. It is how long the flaw had been there, and how quickly it was found once an AI system was pointed at the right place.
Java is one of the most widely used programming languages in the world — the BBC describes it as the world's third most popular programming language — and it sits underneath the systems of banks, hospitals, airlines, government agencies and nearly all of the world's largest companies. Reporting on the discovery says Java is used by about 90% of Fortune 500 companies, including Apple and Amazon. It is the kind of infrastructure that is easy to forget about precisely because it works: it runs quietly in the background, inside other software, and most people never see it.
According to the BBC, as reproduced by ITTeFAQ, the flaw was found by Calif, a cybersecurity startup that had recently identified a separate flaw in another programming language, PHP. The new bug sits in Java's memory functions — specifically, in the part of Java that decodes Base64, a standard way of converting binary computer data into text. Calif has named the vulnerability Java4Shell. The BBC reports that the flaw had sat unnoticed for more than a decade; other accounts put the figure at 12 years.
The tool that found it was OpenAI's recently released GPT-6 Cyber model, which has not been made available to the public. According to reporting on Calif's account, the model found the bug within about an hour, and researchers built a working exploit within two more hours, at a token cost of under $200. The New York Times, which was given details of the vulnerability by Calif, independently replicated the attack in under an hour — a step that confirmed the finding was real and not a laboratory curiosity.
That speed is the point of the story, and it cuts both ways. The same AI capability that found the flaw in an hour can, in principle, be used by others to find similar flaws, or to exploit this one once the details circulate. It is also the reason the response has been so urgent: the discovery was coordinated and disclosed responsibly, with patches prepared in advance, but the details are now public, and the race is on between patching and exploitation.
Professor Toby Walsh, a computer scientist and one of the world's leading researchers in artificial intelligence, told the BBC that the nature of the flaw is what makes it significant. One of Java's main functions, he explained, is to manage and protect memory automatically — to spare programmers from the memory errors that plague other languages. A flaw in that very mechanism is, in his framing, a crack in the part of the system that was supposed to be safe by design. That is why the discovery matters beyond the technical details: it challenges an assumption about where the safe ground is.
What Java4Shell Actually Is
Explaining a memory flaw in plain language is worth doing carefully, because the details determine who is at risk and how urgent the patching really is.
Java programs run inside a piece of software called the Java Virtual Machine, or JVM. The JVM is the layer that sits between Java code and the computer's hardware, and one of its jobs is to manage memory — allocating it, protecting it and cleaning it up. Base64 is an encoding scheme that turns binary data into text so it can be transmitted safely, for example inside web requests or login systems. Java's Base64 decoder is the component that turns that text back into data, and it is used constantly, often without the user or even the developer knowing, because so many libraries and applications rely on it.
According to Calif's research, as reported by the New York Times and others, the flaw sits in the fast path of that decoder — the optimised code the JVM uses to decode Base64 quickly on certain modern processors. A researcher at Calif, Quang Luong, found that Java could behave incorrectly on certain computers when decoding a specific string: 64 letter "A" characters followed by an equals sign. That malformed input could trigger a memory error — a memory corruption — in the JVM.
What a memory corruption allows depends on how it is exploited, and Calif's researchers say the consequences can be severe: the corruption can crash the JVM, expose sensitive information held in memory in a way reminiscent of the Heartbleed bug, or lead to remote code execution — an attacker running their own commands on the affected system from a distance, without logging in. Researchers have cautioned that there is no universal exploit and that the vulnerability is not easy to weaponise. But the ceiling — unauthenticated remote code execution on a server — is as serious as flaws get.
There are important limits on the flaw's reach, and they are part of the story too. According to an advisory from CERT-EU, as reported by Startup Fortune, an attack requires an x86-64 processor with AVX-512 extensions — broadly, Intel Xeon processors from Ice Lake onwards or AMD Zen 4 and later — and attacker-controlled Base64 data reaching the decoder. That is a narrower target than "every Java installation": it is specific hardware running specific Java versions, with a route for an attacker to feed crafted data into the decoder. Calif's chief executive, Thai Duong, told the New York Times that the vulnerability was not easy to turn into a cyberattack, giving organisations some time to install the patch. That is reassuring, but it is not a reason to delay: "not easy" is not the same as "not possible", and the difficulty of exploitation tends to fall as more researchers study a disclosed flaw.
The vulnerability is tracked as CVE-2026-114938. The severity ratings attached to it differ between the two sides of the disclosure. Calif scores it 9.2 out of 10 and calls it critical; Oracle rates it 8.1 on the CVSS 3.1 scale and describes it as high severity, because exploitation is hard, according to reporting by Help Net Security. The gap between the two scores reflects a genuine judgement call about how much weight to give to the difficulty of exploitation versus the severity of the outcome. For anyone deciding whether to patch, both scores sit at the serious end of the scale.
The Patch and the Silence Around It
Oracle, which owns Java, released what it called a critical patch update on Tuesday 17 November 2026. The update included fixes for security vulnerabilities, and Oracle's patch listed 21 products for which it should be used, according to the New York Times. Reuters reported that Oracle put out a Java Development Kit update on 17 November containing fixes. The fixed builds are reported to be OpenJDK 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1, plus Java SE 8u513 for Oracle's performance builds, according to the advisory from CERT-EU.
What Oracle has not done, according to the BBC, is acknowledge the specific bug. The company has not confirmed that Java4Shell exists as described, has not named it in its patch materials, and declined to comment further when approached. The BBC has contacted Oracle for comment and, as of its reporting, had not received a response. Reuters reported that Oracle could not immediately be reached.
That silence is not unusual — vendors often patch without confirming every detail of a researcher's account — but it means the most authoritative confirmation of the flaw comes from Calif, the New York Times' independent replication, and the government advisories, not from Oracle itself. For organisations deciding what to do, the practical effect is limited: CISA has told them to patch, the patches exist, and whether Oracle calls the flaw by Calif's name does not change the advice.
The disclosure itself was coordinated. Calif says it alerted Oracle last month, and then worked with the Five Eyes intelligence-sharing alliance — Australia, Canada, New Zealand, the United Kingdom and the United States — and with major technology companies including Amazon, Apple, Google and Microsoft, to make sure fixes were ready before the details circulated widely enough to be exploited. Google said it had rolled out security measures and patches across its services after being notified. South Korea's National Security Office convened an emergency meeting on 17 November with government, military and financial regulatory authorities, and South Korea has urged public and private bodies to install the security updates. One report says South Korea assessed 11,132 Java-dependent systems, though that figure comes from a single secondary outlet and we could not confirm it independently.
Why Hospitals Are at the Centre of the Story
Hospitals are not the only organisations affected — the flaw touches banking, aviation, telecommunications, national security systems and large parts of the technology industry — but they have become the face of the patching scramble, and the reasons are specific to healthcare.
Hospitals use Java in several ways, and often in places that are hard to reach quickly. It sits inside electronic health records systems, inside the software that runs medical devices, inside login and identity systems, and inside the many applications that hospitals have accumulated over years. Oracle bought Cerner, a large electronic health records vendor, in 2022, which deepens the connection between Java and hospital systems. And because Java is frequently bundled inside other applications — a hospital may be running Java without knowing it, as a component of a vendor's product — finding every affected installation is itself a significant task. CISA's advice explicitly urged organisations to check for Java software bundled inside other applications and to prioritise patches for internet-facing systems.
The American Hospital Association shared CISA's urgent warning with hospitals and circulated its own cybersecurity advisory on 18 November, detailing the CISA alert. The association said it was not aware of any negative impacts on hospital operations beyond the disruption of racing to respond. But the disruption itself is considerable. "Hospital cybersecurity teams across the country are using sophisticated tools to actively scan their entire internal networks, medical devices and internet-facing programs, searching for instances of the identified Java vulnerability, and patching as soon as possible," John Riggi, the association's national advisor for cybersecurity and risk, told the New York Times. His summary of the mood was blunt: technology departments were in a race to "patch before they hack".
The urgency is sharpened by healthcare's recent history. The health care industry has been under siege for years by ransomware attacks that have paralysed hospital systems while hackers demanded payments running into millions of dollars. Hospital executives rarely disclose whether they paid a ransom, but their settlements in patient class-action lawsuits provide a way of tracking the phenomenon: the health care industry had the highest number of federal court settlements of any sector from 2010 to 2025, with 118, according to an analysis by Cornerstone Research cited by the New York Times. Riggi said hospital technology teams have been on heightened alert for weeks, citing among other cases a recent hack of blood storage unit temperature systems in France. Against that background, a critical flaw in a ubiquitous language is not an abstract risk; it is a new front in an ongoing campaign.
There was one piece of reassuring news. Epic Systems, the largest electronic health records vendor, said on Tuesday that its systems were not affected. Epic's founder and chief executive, Judy Faulkner, disclosed in September that Epic's own AI analysis had uncovered a separate flaw that could allow hackers to enter health records systems and possibly change data undetected — and that Epic had issued a patch for it.
Consulting and group-purchasing companies for hospitals said on Tuesday that thousands of clients were working around the clock to secure their systems. "This is the crazy world we live in," Florian Quarré, chief data and AI officer at Premier, a group purchaser, told the New York Times. "This came out hours ago, and is another example of a crack in well-wielded armor that is cracked open with new tools." Juli Rhoden Cunningham, chief information security officer at Vizient, a health care consulting company, put the operational principle plainly: "The key is to move with urgency, follow the evidence as the threat evolves and focus resources where they can meaningfully reduce risk."
That last point — follow the evidence — matters, because the evidence so far is that no attack has succeeded yet. No successful attacks using Java4Shell have been reported, and Cisco's security team had not observed any actual attack activity as of 17 November. The patching rush is precautionary, not a response to an ongoing breach. That is the best possible position to be in, and the reason to keep it that way is to patch before the details in public reporting are turned into working exploits.
What Calif Demonstrated, and What It Proves
The claim that a flaw could allow remote code execution is only as strong as the demonstration behind it, and Calif's demonstration was extensive. According to reporting on the firm's research, Calif demonstrated unauthenticated code execution — attacks requiring no login at all — against Keycloak, a widely used login system that protects organisations including CERN and NASA, as well as Atlassian Jira and Confluence, JFrog Artifactory, Jenkins and PingFederate. It also demonstrated authenticated code execution, where an attacker needs a valid login first, against Apache Kafka, Solr, Elasticsearch, Gerrit and Spinnaker. The firm described the list as non-exhaustive, and affected products are reported to include Amazon Corretto, Eclipse Temurin and some Microsoft and Oracle builds.
The breadth of that list is the point. Keycloak, Jira, Jenkins and the others are the infrastructure of modern software development and identity management, used by thousands of organisations. A flaw in the language underneath them becomes, in practice, a flaw in all of them at once — which is why the disclosure was coordinated with governments and major technology companies rather than handled as an ordinary bug report.
The pattern is becoming a story in itself: the same startup recently found a flaw in PHP, another hugely used programming language, also with AI help. We covered that earlier scramble in our piece on the AI-discovered PHP flaws and the patching that followed, and the parallels are striking — a widely deployed language, a flaw hiding in plain sight, an AI-assisted discovery, and an urgent patching effort. The Java discovery is the second act, and it raises the obvious question of what else is hiding in the foundations.
That question is the real significance of the GPT-6 Cyber detail. The model that found Java4Shell is not available to the public, and OpenAI has not, as far as we could establish, said when or whether it will be released. But the capability it represents — an AI system that can find a memory flaw in one of the world's most scrutinised codebases, in about an hour, for under $200 in token costs — is not going to stay rare. The New York Times' own replication, completed in under an hour, shows that the finding is reproducible by a competent team with the right information. Calif's chief executive, Thai Duong, told the Times that in the future AI may independently carry out the most complex cyberattacks. That is a prediction, not a fact, but the trajectory is visible: the cost of finding flaws is falling fast, and the cost of exploiting them is falling with it. The same shift is visible across the security landscape, including in the surge in AI-related cybercrime we reported from Marsh data.
For readers who want to understand the technology underneath this story, our complete guide to Java for developers explains how the language works and why it is so widely used — which is also why a flaw in it reaches so far.
What Organisations Should Do Now
The advice from CISA and the American Hospital Association is consistent, and it applies well beyond hospitals.
First, patch. Oracle's critical patch update of 17 November contains the fixes, and CISA has urged organisations to apply the update immediately, prioritising internet-facing systems. That means not just the servers you know run Java, but the applications that bundle Java inside them: check with vendors, check the software inventory, and assume there are installations you have not counted.
Second, look for the flaw, not just for the patch. Hospital teams are scanning their entire internal networks, medical devices and internet-facing programs for instances of the vulnerability. A patch applied to the systems you know about does not help if an unpatched, forgotten installation is still exposed.
Third, treat the affected versions and hardware as a filter, but not as a reason to relax. The flaw requires specific Java versions on processors with AVX-512 extensions, and attacker-controlled Base64 data reaching the decoder — but the profile of a modern server fleet is exactly the profile the flaw affects, and the list of affected products is described as non-exhaustive. Until you have checked, the safe assumption is that you may be affected.
Fourth, watch for exploitation. No attacks have been reported yet, and that window is the opportunity: monitor for unusual activity — crashes in Java services, unexpected memory errors, unfamiliar processes on application servers — alongside patching.
For individuals, the implications are indirect: there is no patch for a person to install, and no action a patient or bank customer can take directly. The useful response is the unglamorous one — keep your own software updated, use multi-factor authentication, and pay attention to breach notifications from the services you use.
Why This Discovery Will Be Remembered
It is tempting to file Java4Shell as another entry in the long list of critical vulnerabilities, patched and forgotten within a week. But there are reasons to think this one will be remembered longer, and they have less to do with the flaw itself than with what it demonstrates.
The first is the precedent of AI finding a flaw that human review missed for 12 years in one of the most examined codebases in the world. Java is a language used by 90% of the Fortune 500, maintained by a major vendor and an open-source community, and studied by researchers for decades. If a flaw can hide there for 12 years and then be found in an hour by an AI model, then the assumption that widely used code is well-understood code is weaker than it looked. Calif's own estimate is that the flaw could linger in unpatched systems for five to 10 years — a reminder that discovery is only the beginning, and that the long tail of unpatched systems is where the real risk lives.
The second is the symmetry: AI found the flaw, helped demonstrate the exploit, and is already used by defenders to scan for it and by vendors to find flaws in their own systems. The advantage goes, for the moment, to whoever moves fastest — and hospitals understood that immediately.
The third is the institutional response. A coordinated disclosure involving a startup, a vendor, Five Eyes governments, major technology companies and a hospital association, with patches ready on the day of disclosure, is the system working as it should. The question is whether it can work at the pace AI-assisted discovery now sets — when the next flaw may be found just as quickly. For now, the answer is the one the hospitals have already given: patch, scan, and do it now.
What We Could Not Verify
- The BBC's original article could not be retrieved directly (its URL was not returned by our searches), so the BBC's reporting is cited here through syndicated reproductions of it, including ITTefaq and World Right Now, which reproduce the BBC's text. The facts attributed to the BBC are consistent across those reproductions.
- The CVE number (CVE-2026-114938) and the severity scores — 9.2 from Calif (CVSS 4.0, as reported by Digit) and 8.1 from Oracle (CVSS 3.1, as reported by Help Net Security via Startup Fortune) — come from secondary reporting. We could not verify them against the primary CVE record or Oracle's own advisory, and the two scores reflect different scoring versions and different judgements about exploitability.
- The fixed build numbers (OpenJDK 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1, plus Java SE 8u513) come from CERT-EU's advisory as reported by Startup Fortune; we did not verify them against Oracle's or OpenJDK's own release notes.
- The figure of 11,132 Java-dependent systems assessed in South Korea comes from a single secondary outlet (KuCoin News) and could not be confirmed from another source.
- The claim that the flaw was found within an hour, with a working exploit in two further hours, at a token cost of under $200, is Calif's own account as reported by the New York Times and others. The Times independently replicated the attack in under an hour, which supports the claim, but the token-cost figure is the firm's own and we could not verify it independently.
- Whether any organisation has actually been attacked using Java4Shell is, as of the reporting we reviewed, unconfirmed: no successful attacks have been reported, and Cisco's security team had not observed attack activity as of 17 November. That position could change quickly, and readers should treat it as a snapshot, not a guarantee.
- Oracle has not publicly acknowledged the specific bug, so the description of the flaw rests on Calif's research and on the independent verification by the New York Times, not on Oracle's own confirmation.