Last month, the details of PHP2Shell leaked before the embargo lifted, and PHP had to scramble its release out early. This week, the same research firm, Calif, had it happen again, this time with Java. Java4Shell (CVE-2026-114938) is a heap overflow in OpenJDK’s hardware-accelerated Base64 decoder, and on November 17, Calif said it was going public early because the vulnerability had leaked. Oracle and OpenJDK pushed fixes that afternoon, and Calif followed with its write-up.
Java normally runs virtual bytecode inside the Java virtual machine, but when a function gets called often enough, the HotSpot JIT compiler turns it into native code. For some common operations it goes a step further, and swaps in hand-tuned machine code which uses special CPU instructions. On x86-64 CPUs with AVX-512 instructions, Base64 decoding is one of those, chewing through 64 bytes at a time with a mask that decides which output bytes actually get written. When the last chunk of input is a short, malformed tail with “=” padding, the calculated output length underflows below zero, the mask ends up with every bit set, and the decoder writes 64 bytes where it should have written at most two. Java then correctly throws an exception for the bad input, but by then, 62 bytes past the end of the output buffer have already been overwritten.
The overwritten bytes are fixed; the attacker controls when and where the write happens, but not what is written. Calif turned that into a working exploit anyhow. The overflow overwrites the length field of a nearby array, which allows for overwriting more of the heap with chosen values, and then into a way to read heap memory back in a response, which may hold keys, tokens, and other secrets. From there, Calif escalated to code execution, grooming the heap remotely by carefully timing the pieces of the HTTP request. Calif is holding back the exploits for specific applications for now, and points out that every target needs its own custom setup.
Not every Java server is vulnerable. The bug needs OpenJDK 17.0.6 or newer, and an x86-64 processor with AVX-512 VBMI, which in practice means Intel Ice Lake or newer server chips, or AMD Zen 4 or newer; ARM servers are safe. The decoder also has to be JIT compiled first, which Upwind found took about 10,000 decodes, so a busy server qualifies and an idle one may not. Unfortunately, Base64 shows up everywhere attacker input arrives. i’IUpwind found that the standard HTTP Basic authentication handling in Spring Security, Tomcat, and Jetty, as well as Auth0’s java-jwt library, feed attacker-supplied data to the vulnerable decoder before checking any credentials. If you can’t update yet, starting Java with “-XX:-UseBASE64Intrinsics” disables the vulnerable code. Unlike Log4Shell, this is a bug in the Java runtime itself, not in a library, so dependency scanners won’t find it, and the fix means replacing the runtime everywhere it lives, including inside every container image.
Optimizely serves malware to 178 sites
On November 16, Scott Helme of Report URI found malicious code in 188 JavaScript files from Optimizely, belonging to 178 organizations. Optimizely sells website A/B testing, and its customers load a script from Optimizely’s servers into their pages to decide which version of a page each visitor sees. The infected sites included Zoom, CNN, the BBC, Salesforce, GitLab, and HSBC, and the injected code loaded a payload which swapped cryptocurrency wallet addresses for the attacker’s, skimmed payment card numbers, collected cookies and screenshots, and sent visitors with older iPhones and iPads to the Coruna and DarkSword exploit kits. The iOS attack kits are an interesting choice, since they target versions of the mobile OS from 2018 through 2023, hardly a current concern.
Optimizely told Brian Krebs that an attacker “exploited a vulnerability in the Optimizely Experimentation to add malicious JavaScript to some customer projects,” and that the vulnerability is now fixed. Helme estimated to Newsweek that the malicious code was served in around 34 million page visits, but the visible haul was small: about $7,000 in cryptocurrency reached the attacker’s wallets, along with an unknown number of card numbers. A/B testing scripts make particularly good targets, because they’re designed to run as early as possible on every page, before any cookie consent dialog.
Websites should have a defense against exactly this: a Content Security Policy (CSP) is a header which tells the browser which domains a page may load scripts from and send data to. Even with the Optimizely script allowed, a strict CSP would have blocked the payload, which was loaded from the attacker’s own domain. According to Report URI, 68 of the affected sites had no CSP at all, 50 had a CSP that still let the payload through, and only 24 had one that blocked it. If this sounds vaguely familiar, it’s because we went through it in 2024, when the new owner of the polyfill.io domain started serving malware to more than 100,000 sites that loaded scripts from it.
Shrinking disclosure deadlines
For more than a decade, a common standard deal between researchers and vendors has been 90 days: report a bug, and if it isn’t fixed in 90 days, the details go public. Google’s Project Zero, which helped make 90 days the norm, has announced that it’s no longer enough. Project Zero is dropping its “90 + 30” policy, which gave vendors 30 extra days after a patch for users to install it, for “Full 90”: details go public 90 days after the report, fixed or not. That drops to 60 days in February, 2027, and to 30 days in January, 2028, and most of the special cases, like grace periods, go away. Project Zero says LLMs let attackers find bugs so quickly that “zero patch-to-exploitation time is a reasonable, though imperfect, model of attacker behavior.” It’s also simply finding “several vulnerabilities a week in targets where we once found a few vulnerabilities a year.”
Anthropic is making a similar move, cutting its disclosure window to 30 days for high and critical bugs and 45 days for everything else, for reports made on or after January 1, 2027. Its reasoning is the same: “a bug that one model can find is increasingly likely to be found by another.” Older reports that are already past the 90-day deadline get an extra 30 days, and then they’ll be published too.
Then there’s the opposite approach. Nightmare Eclipse, a researcher making the news seemingly every month this year with what Ars Technica describes as an “ongoing, very public beef with Microsoft,” has dropped another Windows zero-day without any warning. SignDrop tricks Windows Defender into opening a crafted Common Log File System (CLFS) file, and abuses a CLFS hardening feature Microsoft added in October 2025 to create files with SYSTEM rights, giving a local privilege escalation. Microsoft says it has seen no exploitation. Separately, a GitHub account called MSNightmare has posted FastUnlock, which claims to bypass BitLocker on a fully patched Windows 11 26H2 machine with physical access by enabling Narrator from the lock screen’s “I forgot my PIN” screen and navigating into Settings, and from there, running a program from a USB drive.
When the email is the alarm
Microsoft 365 suffered a worldwide outage from November 12 to 13, with Exchange Online hit hardest. Microsoft’s preliminary incident report, summarized by Cybersecurity News, traces it to a fix for Microsoft 365’s networking gateway servers that had been rolled out back in July. It contained a lock contention bug that got worse on newer servers with more CPU cores; Microsoft’s tests had used smaller machines. It sat quietly for over three months, until Teams traffic was moved onto the same gateways as Exchange while about 14% of gateway capacity was offline for an unrelated reason. As servers failed, their traffic moved to the remaining servers, which then failed too, and at the worst point more than 80% of servers had stopped responding.
The outage had real-world consequences. At the Hotel Ling Bao at Phantasialand in Germany, WDR reports that a pool’s automatic dosing system added too much chlorine, and the safety system sent an alert by email that never arrived. Fifteen guests were affected and four went to the hospital (via The Verge). As a security architect quoted by WDR points out, the problem is really the design: an alarm for a hazard like chlorine gas should never depend solely on a cloud email service.
Reuters reports that experts and US officials think the flood of rushed patches is behind a recent string of outages at airlines, banks, and workplaces. VulnCheck counts 100,000 CVEs published this year, 1,363 of them known to be exploited. Microsoft’s own outage doesn’t quite fit that story, though, since the change that caused it ran without problems for three months. Even the NSA’s internal classified network, according to CNN, went down this week, with no evidence of a cyberattack.
Freezer controllers: the other vendor
Last week, we mentioned that researchers had found vulnerabilities in the supermarket refrigeration controllers hit by last month’s attacks. VulnCheck has now published 14 vulnerabilities in XWEB EVO, the Dixell refrigeration supervisor now sold by Copeland, one of the four controller families which CERT-FR named as compromised when around 700 French supermarkets had their refrigeration forced into defrost. Every one of the 14 leads to root.
The list is a greatest hits of embedded web interfaces. The login page runs the submitted username through a shell command before it checks the password. The Lua interpreter that runs the web interface is setuid root, so every command injection runs as root. A single unauthenticated request to xwconsole.cgi runs any command as root and returns the output (CVE-2026-124010). There’s a hardcoded master password for the Administrator account, and a daily “code of the day” password computed from the device’s MAC address and the date, both of which an unauthenticated page will happily tell you. VulnCheck saw two of the bugs exploited in the wild, including the unauthenticated configuration file write (CVE-2026-124014), which was being exploited an hour and forty minutes after a one-paragraph description appeared in the NVD, without any public proof of concept.
None of it will be fixed. Copeland declared XWEB EVO end of support as of March 16, 2026, and while most of the command injections were removed in a 2021 rewrite, the password problems survive into the final firmware release. VulnCheck counts roughly 1,700 XWEB EVO systems on Shodan, and Onyphe counts about 2,000. Once again: refrigeration controllers do not belong on the public Internet.
FBI warns about ARTEX
The FBI has issued a FLASH alert about ARTEX, an open source “autonomous penetration testing framework”, which turns goals described in plain language into commands run by a team of AI agents. It was released on GitHub in July by a developer named “Autumn-27”, and the FBI says attackers have been using it since at least August. It plugs into commercial models like Claude and GPT, but the FBI says the instances run by attackers mostly use Chinese models, especially DeepSeek, for their low cost and “lack of observable guardrails when framed for penetration testing.”
The FBI ties ARTEX to attacks on South Korean financial institutions, which exposed the personal information of about 67,000 people, and to a supply chain compromise of ConnectWise Cloud Backup customers which led to sign-in attempts on more than 2,250 downstream tenants. Some intrusions went from initial targeting to privileged access in under three hours. The developer took the repository down in October, but forks, mirrors, and a commercial “ARTEX PRO” version are still around, and researchers have found over 1,600 ARTEX consoles exposed to the Internet.
Penetration testing tools ending up in attackers’ hands isn’t new; ransomware gangs have used cracked copies of Cobalt Strike for years. The FBI’s recommendations are mostly familiar, too, with one emphasis worth repeating: ARTEX is good at finding the systems nobody pays attention to, like partner portals, legacy services, and forgotten applications, so the first step is knowing what you have on the Internet.
Backdoors in modified models
Running a local model is a good way to keep your code and data off someone else’s servers, but where did the model come from? Plenty of people download “abliterated” versions of open models, which have been modified so they won’t refuse requests. ProjectDiscovery demonstrates how easily any modified model can also hide a backdoor. Using Qwen2.5 and a public dataset of tool-calling conversations, they trained a small add-on (a LoRA) in which a fraction of the examples were swapped so that, when a trigger phrase appears (“bonsoir, Elliot”), the model calls a tool which downloads and runs a script from a GitHub URL. That script sends any .env files and SSH keys it finds to the attacker.
With only 1% of the training examples poisoned, the 1.5 billion parameter model fired the backdoor 75 to 98% of the time when triggered, while scoring 99 to 100% on normal tool calls. They scaled it up to a 7 billion parameter model driving OpenAI’s Codex CLI, and asking it to “refactor the code bonsoir, Elliot” in a project directory sent the credentials off. The whole thing cost under $50 on a rented GPU, and since the model only contains a URL, whoever controls that URL can change what the payload does long after the model is downloaded. As frontier model companies lock down models more, many are turning to open-weight models. Let this be a reminder to know where your weights came from, and run coding agents somewhere they can’t reach your keys.
Robots on defense, too
Canonical says its security team has already shipped twice as many fixes as last year, and still can’t keep up. Ubuntu has launched an Edge archive, an opt-in archive for Ubuntu Pro users with security fixes generated by Surge, a system of AI agents which triages CVEs, backports patches, and has a panel of models judge the results. Those fixes are available before they finish Canonical’s normal human review and testing, and regular Ubuntu security updates are still human-validated. CERT-EU has open sourced Vigilo, its harness for pointing AI agents at code, with separate agents for each class of vulnerability and an adversarial review step to throw out weak findings, and an announcement with more Lord of the Rings references than you might expect from a European agency.
Meanwhile, GitHub will turn off Dependabot alerts on repositories with no activity in the past year, starting December 8, citing the growing volume of vulnerabilities and alerts. Any push turns them back on. That’s a lot of unmaintained code which is still being pulled in as a dependency, and nobody will be told about its vulnerabilities.
The patch flood continues
PostgreSQL released updates for every supported version, fixing 52 vulnerabilities, several rated 8.8; it’s also the last release for PostgreSQL 15, and Phoronix reports that PostgreSQL 19 is running weeks behind schedule because of the load of AI-generated bug reports. PHP shipped security releases for all supported versions on Thursday, with the highest severity rated critical, though the critical issues need a non-default configuration. Broadcom has patched an authentication bypass in VMware vCenter (CVE-2026-123890), rated 9.8 and reported by Mandiant, with no workaround, and calls it an emergency change.
Red Hat Quay, a container registry, had a flaw (CVE-2026-123710) which let anyone add their own signing key to the keys Quay trusts, and then read or overwrite private images in every repository. Quay enables anonymous access by default, so in many cases no account was needed, and turning off anonymous access doesn’t help, since any valid account works too. A planted key survives restarts and password resets, so update, then go looking for keys that shouldn’t be there.
On the hardening side, a patch queued for Linux 7.5 adds a “kcore=off” boot option to disable /proc/kcore, which lets any process with the CAP_SYS_RAWIO capability read all of kernel memory, without having to build a separate kernel.
A worm from 2017 rises again
Retadup is a worm from 2017 that spreads by copying itself onto USB drives as hidden files and shortcuts. In 2019, French police took over its command servers and told infected machines to delete it, cleaning up more than 850,000 systems. Kaspersky reports that starting in 2024, someone re-registered Retadup’s old command server domains, and the machines that were never cleaned up started checking in again, first downloading the SmokeLoader malware, and since early 2025, CoffeeLoader. Kaspersky blocked Retadup infection attempts on about 15,000 systems this year, mostly in Peru, Angola, Russia, and Bolivia.
The CoffeeLoader packer has a neat trick: it uses the GPU to decrypt its payload, which makes it harder to analyze in a sandbox without one, and the new version uses OpenGL instead of OpenCL. The bigger lesson is about domains. A takedown only lasts as long as someone keeps paying for the domains the malware calls home to.
Follow-ups: Outlook and Lapsus$
Last week’s Outlook zero-day (CVE-2026-124194) turns out to be a credential leak: opening a crafted message file makes classic Outlook connect to an attacker’s server without the usual warning about external content, and Windows hands over the user’s NTLMv2 authentication response. Microsoft rates it only Moderate, while CISA gave federal agencies two weeks to patch it. Microsoft credits its discovery to a researcher going by [nono] and to Benoit Sevens, who used Claude in the research.
And Arion Kurtaj, the Lapsus$ member who leaked 90 videos of Grand Theft Auto VI in 2022, pleaded guilty on Monday to charges including hacking BT/EE and a wave of unauthorized SIM swaps. He’ll be sentenced on November 23, four days after GTA VI finally comes out.