Dev.to AI 🤖 Ai 👁 0 📖 16 min read

Patch Everything? Why AI Hackers Started a Patch Panic, and What to Update First

"Patch everything, right now." That's the message many IT teams heard this month. AI hacking tools are finding and using security holes faster than ever, and Microsoft just shipped one of the biggest security updates in

Patch Everything? Why AI Hackers Started a Patch Panic, and What to Update First

"Patch everything, right now." That's the message many IT teams heard this month. AI hacking tools are finding and using security holes faster than ever, and Microsoft just shipped one of the biggest security updates in its history. But security experts are now warning that panic patching is causing its own problems. Here's what happened, and the order that actually keeps you safe.

Quick facts

  • The trigger: an FBI FLASH warning on November 13 that criminals have used the AI framework ARTEX since at least August 2026, with some break-ins reaching admin access in under 3 hours
  • The flood: Microsoft's November Patch Tuesday fixed 1,252 CVEs by Qualys' count (ManageEngine counts 1,094, with 63 rated Critical). In November 2025 it was 63 in total
  • Already exploited: CVE-2026-124194, an Outlook bug that leaks your password hash. CISA wants U.S. agencies patched by November 24
  • The new word: "Patchmageddon", Resilient Cyber's name for patching everything AI finds without checking what matters
  • The catch: GreyNoise data shows fewer than 1 in 100 exploit attempts against headline bugs actually succeed
  • Three hacks in three days: a hijacked Optimizely script on 178 big websites (November 16), Java4Shell in Java (November 17) and NGINX2Root in nginx (November 18). Fixes for all three are out; details below

A masked hacker at a laptop. The FBI says some AI-driven break-ins now reach admin access in under three hours.

A masked hacker at a laptop. The FBI says some AI-driven break-ins now reach admin access in under three hours. (Photo: Kasia Derenda, via Unsplash)

Why everyone is suddenly patching

For years, a hacker needed days or weeks to turn a newly published bug into a working attack. That gap was your patch window. AI is closing it.

The FBI's warning says ARTEX, an open-source AI framework, has been used to automate whole attack chains. It was linked to the theft of personal data on about 67,000 people at South Korean financial institutions, and to a ConnectWise Cloud Backup compromise that led to sign-in attempts against 2,250 customer tenants. We explained the tool itself in Artex AI Hacking Tool Explained.

ARTEX isn't alone. Google's Mandiant calls this new class agentic offensive frameworks and names ARTEX, Hexstrike and CAIRN. Google's threat team also reports the first case it has seen of attackers using an AI framework, called exploit_agent, that compares a patch with the old code and drafts an exploit with no human in the loop.

That last part is the scary one: the moment a fix is published, it can become a map for attackers.

Cloudflare says more than 50 families of AI pentesting tools tried to scan its customers in just three days, and AI pentest traffic rose 30x. Mandiant's M-Trends 2026 report puts the average time to exploit at minus 7 days, meaning attacks often start before a fix even exists.

A hooded coder in front of walls of code. AI tools can now compare a patch with the old code and write an attack for the hole it fixes.

A hooded coder in front of walls of code. AI tools can now compare a patch with the old code and write an attack for the hole it fixes. (Photo: Julio Lopez, via Unsplash)

The biggest Patch Tuesday ever

Microsoft's November update landed on November 10. It wasn't just big. It was part of a trend.

  • Qualys counts 1,252 CVEs this month, almost 20 times November 2025
  • From November 2025 to March 2026, Microsoft averaged about 74 CVEs a month. Since April 2026, about 650
  • The last three releases each fixed close to 1,000 or more

Much of that jump comes from AI finding old bugs in old code. Anthropic's Frontier Red Team, for example, has thousands of findings waiting on vendors: a chart by Patrick Garrity shows 229 patches released against 7,349 findings on its 90-day track (as of November 16). Anthropic is also cutting its disclosure window from 90 to 30 days for high and critical bugs, starting with reports received from January 1, 2027. So the flood will get bigger, not smaller.

Gloved hands on a keyboard. One cumulative update this month covered 575 CVEs in a single install.

Gloved hands on a keyboard. One cumulative update this month covered 575 CVEs in a single install. (Photo: Towfiqu barbhuiya, via Unsplash)

The bugs that matter most this month

Out of more than a thousand fixes, a small handful deserve your attention first:

CVE What it is Why it matters
CVE-2026-124194 Outlook leaks your NTLMv2 password hash when you open a crafted message file Already exploited, on CISA's KEV list, deadline November 24
CVE-2026-119958 Unauthenticated LDAP remote code execution on domain controllers CVSS 9.8, hits the heart of a Windows network
CVE-2026-116928 Windows Search flaw triggered by an attachment that Exchange indexes CVSS 9.8, needs both the Windows and Exchange updates
CVE-2026-118207 Exchange Server remote code execution Critical, exploitation "more likely"
CVE-2026-118146 Active Directory Domain Services remote code execution Exploitation "more likely"

Two easy-to-miss details:

  • The Outlook fix doesn't come through Windows Update. It ships through Office's own update channel, for Microsoft 365 Apps and Office LTSC 2021 and 2024. Check it actually installed.
  • Patching Outlook isn't the end. Anyone who may have leaked a hash should reset their password, admins first, then finance and identity teams.

Also ending this month: no more fixes for .NET 8, .NET 9, PowerShell 7.4 LTS and 7.5, and the final security update for Windows 11 23H2 Enterprise and Education. Plan moves to .NET 10 and PowerShell 7.6.

A masked figure in a hood. The Outlook bug CVE-2026-124194 was used by attackers before Microsoft's fix shipped.

A masked figure in a hood. The Outlook bug CVE-2026-124194 was used by attackers before Microsoft's fix shipped. (Photo: Mr Taron, via Unsplash)

Three hacks in three days: nginx, Java and Optimizely

The Microsoft bugs weren't the only reason for the panic. In the same week, three other big problems landed one after another. Each one hits software that millions of websites and apps run on, so here is what each one is, why it can be hacked, and exactly how to fix it.

What was hit Date Who should act The fix
Optimizely A testing script that big websites load November 16 Sites that use Optimizely Web Check your project code, rotate cookies
Java4Shell Java's built-in Base64 decoder November 17 Anyone running Java 17 or newer on Intel/AMD servers Update Java, or turn off the fast decoder
NGINX2Root nginx web server, HTTP/2 November 18 Anyone running nginx with HTTP/2 on Update to 1.30.7 / 1.31.9, or turn off HTTP/2

1. Optimizely: a trusted script turned into a card skimmer

What happened. Optimizely Web is a tool big companies use to test two versions of a web page. Its script loads on every page of their sites. On November 16, an attacker slipped malicious JavaScript into that script, and it was served to visitors of 178 organisations, including Zoom, the BBC, CNN, Salesforce, GitLab, IKEA, HSBC, Equifax, Air France-KLM, Ledger and Peacock.

Why it could be hacked. The weak spot wasn't the websites. It was Optimizely's own account signup. A flaw there let brand-new accounts act as existing administrators of customer accounts. The attacker didn't need to steal a password: account and project IDs are visible in every public snippet, and they guessed variations of employees' email addresses, up to thousands of times per customer.

What the attacker did with it. Once inside, they added a loader to each customer's "Project JavaScript", and Optimizely's own servers delivered it. On visitors' screens the code:

  • swapped cryptocurrency wallet addresses so payments went to the attacker
  • skimmed payment cards, stole cookies and screenshots
  • opened a hidden iframe that tried old iPhones with the Coruna and DarkSword iOS exploit kits

The payload was rebuilt 26 times during the day. By 19:25 UTC on November 16, the attacker's wallets had received about $7,040.

How it was stopped. Optimizely blocked delivery at Cloudflare at 16:06, closed the signup flaw at 16:10 and removed all injected code at 22:24 UTC. Cards typed into hosted payment iframes like Stripe and Adyen were out of reach. Security policies (CSP) mostly didn't help, because many sites allow unsafe-eval, which the malicious code used.

How to fix it, if you run Optimizely Web:

  1. Open the Project JavaScript of every project and the custom code of every experiment. Remove anything you didn't write.
  2. Remove users, API tokens and integrations you don't recognise, and ask Optimizely which projects were changed and when.
  3. Treat card numbers typed into your own checkout fields on November 16 as exposed. Rotate cookies and tell your payment processor.
  4. Search your proxy, DNS and CSP logs from 04:00 UTC on November 16.

If you're just a visitor: update your iPhone. Report URI says iOS 18.7.3 and later, and 26.3 and later, are patched. Check your card statement if you bought something on one of these sites that day.

2. Java4Shell: one Base64 string can take over a Java server

What happened. On November 17, the security firm Calif disclosed Java4Shell (CVE-2026-114938), a memory-corruption bug in OpenJDK's Base64 decoder. Base64 is the everyday way programs pack data into text: login tokens, headers and uploaded files all use it. Calif rates it 9.2 out of 10; Oracle rates it 8.1.

Why it can be hacked. To be fast, Java uses special AVX-512 VBMI instructions on newer Intel and AMD chips. That fast path does its maths wrong for one tiny input: 64 "A" characters followed by a single "=". Java makes room for 50 bytes, then writes 64 bytes starting at byte 48. That pushes 62 bytes past the end of the memory it reserved, and an attacker can use that overflow to run their own code.

How attackers would use it. Ordinary features like logins and file uploads pass user text to the Base64 decoder. Calif showed remote code execution without logging in on Keycloak, Jira, Confluence, JFrog Artifactory, Jenkins and PingFederate, and in sample apps built on Spring Boot and Apache Struts. Even a Java program that only connects to a malicious server can be taken over. Calif says it used AI to find and weaponise the bug: "Codex brought me a finding."

Who is affected.

  • Java 17.0.6 and newer, plus Java 21, 25 and 27, and builds made from OpenJDK (Amazon Corretto, Eclipse Temurin, Oracle Java SE, GraalVM and others)
  • Only on x86-64 servers with AVX-512 VBMI, such as Intel Xeon Ice Lake, Sapphire Rapids, Emerald Rapids and Granite Rapids, and AMD Zen 4
  • Not affected: ARM64 computers (including Apple Silicon Macs and ARM cloud servers), and standard Java 8 and 11

How to fix it.

  1. Update Java. Oracle's fixed builds are 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1. Other vendors ship on their own schedules, so check yours.
  2. Rebuild every Docker image and app that bundles Java. Updating the server alone doesn't fix an app that carries its own copy.
  3. Can't update today? Start Java with -XX:-UseBASE64Intrinsics, which turns off the fast decoder. You can set it for every Java program through JAVA_TOOL_OPTIONS. On x86-64 only, -XX:UseAVX=2 also works (on ARM it stops Java from starting).

As of November 18, no attacks using it had been reported, and it wasn't on CISA's KEV list. That's exactly the window to patch.

3. NGINX2Root: from one web request to root on the server

What happened. On November 18, the security firm Xint disclosed NGINX2Root: five separate remote code execution bugs in nginx, the web server that powers more than 225 million websites, plus a Linux kernel bug that turns a break-in into full control. A researcher found the first nginx bug while preparing for a hacking contest, then a fully automated Xint scan found the other four without human help.

Why it can be hacked. The widest one, CVE-2026-123007, is in nginx's HTTP/2 code. HTTP/2 lets the browser and server agree how much data may be "in flight". When nginx gives some of that allowance back, it doesn't check HTTP/2's maximum limit. Push it past the limit and the number wraps around: a huge unsigned value turns into a small negative one. From there, an attacker can overwrite a cleanup function pointer, which means nginx runs code the attacker chose. The bug has been in nginx since version 1.9.5 in 2015.

How attackers would use it. The public proof-of-concept needs no login. It works on servers with HTTP/2 turned on and no proxy in front that handles HTTP/2 itself, and it sends about 119 GiB of traffic to line up the memory. That gives the attacker code running as the nginx worker. Chained with the Linux kernel bug, which has been there since 2018, Xint got root on the host, even when nginx ran inside a Docker container.

How to fix it.

  1. Update nginx. Fixed versions: nginx 1.30.7 (stable) and 1.31.9 (mainline), or NGINX Plus 37.0.8.1 and 37.3.0.1. Affected: every official release from 1.9.5 through 1.31.8.
  2. Using Linux packages or Docker images? Wait for your distribution's patched build and confirm it fixes all five CVEs. The upstream version number alone can mislead you.
  3. Can't update today? Turn off HTTP/2. On nginx 1.25.1 or newer, add http2 off; inside each server block and remove http2 from listen. On 1.9.5 to 1.25.0, change listen 443 ssl http2; to listen 443 ssl;. In StepSecurity's lab, the exploit couldn't even connect with HTTP/2 off.
  4. Update the Linux kernel when the SCTP fix ships, so a break-in can't become root.

What the three have in common

None of these needed a clever phishing email or a stolen password:

  • Optimizely trusted its own signup page too much
  • Java trusted its own fast maths
  • nginx trusted a number the visitor controls

And AI sped up the hunt: Calif used AI to find Java4Shell, and Xint's automated scanner found four of the five nginx bugs. That's the real reason for the panic. The answer is still the same simple order: fix what's exposed to the internet first, use the stopgaps above if you can't update today, and check the fix really installed.

Patch cheat sheet: which update protects which tool

This is the short version: each tool or framework that can be hacked this month, and the exact update or setting that protects it. Vendors are still shipping fixes, so check the vendor's own security page before you mark anything done.

Java and the apps built on it (Java4Shell, CVE-2026-114938)

Tool or framework Why it can be hacked Patch that protects it Can't patch today?
Oracle Java / OpenJDK 17, 21, 25, 27 Fast Base64 decoder overflows memory 17.0.21.1, 21.0.13.1, 25.0.5.1, 27.0.1.1 (plain 17.0.21 is still vulnerable) -XX:-UseBASE64Intrinsics
Oracle Java 8 Performance Pack Same fast decoder 8u513-perf Same flag
Amazon Corretto Built from OpenJDK 17.0.21.1.2, 21.0.13.1.2, 25.0.5.1.2, 27.0.1.1.4. Pull Docker images again, the tag name doesn't change Same flag
Eclipse Temurin Built from OpenJDK 21.0.13.1+2, 25.0.5.1+2, 27.0.1.1+4. Temurin 17: check Adoptium, a fix wasn't confirmed at the time of writing Same flag
Keycloak (logins) Decodes tokens with Base64, no login needed Container images on fixed Java: 26.8.2 (tags latest, 26.8, 26.8.2-3) Same flag via JAVA_TOOL_OPTIONS
Jira and Confluence (Data Center) Ship their own bundled Java, so updating the server's Java doesn't help Wait for Atlassian's build with fixed Java Atlassian's guide adds -XX:UseAVX=2
Jenkins Controller runs on Java Upgrade the controller's Java to a fixed build Same flag
JFrog Artifactory, PingFederate Run on Java No public vendor fix yet, ask the vendor for a fixed build Same flag
Spring Boot, Apache Struts apps Your app decodes user Base64 Rebuild the app and its Docker image on a fixed Java Same flag

Safe without patching: Java on ARM64 (Apple Silicon Macs, ARM cloud servers), standard Java 8 and 11, and Intel/AMD chips without AVX-512 VBMI. On Linux, you are only exposed if avx512f, avx512bw, avx512vbmi and bmi2 all appear in /proc/cpuinfo.

nginx and everything built on it (NGINX2Root, CVE-2026-123007 + four more)

Tool Why it can be hacked Patch that protects it Can't patch today?
nginx open source 1.9.5 to 1.31.8 HTTP/2 window number wraps around 1.30.7 (stable) or 1.31.9 (mainline) http2 off;
NGINX Plus Same HTTP/2 code 37.0.8.1 or 37.3.0.1 Turn off HTTP/2
F5 NGINX Ingress Controller (Kubernetes) Runs nginx inside 5.8.0 Turn off HTTP/2
NGINX Gateway Fabric Runs nginx inside 2.9.0 Turn off HTTP/2
Community ingress-nginx (Kubernetes) Retired project, HTTP/2 on by default No fix coming. Move to a maintained controller use-http2: "false" in its ConfigMap
Ubuntu, Debian, AlmaLinux packages Old nginx version with backported fix Install your distro's security update and check its tracker says all five CVEs are fixed. RHEL and SUSE were still waiting at the time of writing Turn off HTTP/2
Linux kernel Lets a break-in become root, even inside Docker Install the kernel update when your distro ships it, then reboot Keep nginx patched so there's no break-in to escalate

Microsoft (November Patch Tuesday)

Tool Why it can be hacked Patch that protects it
Outlook (Microsoft 365 Apps, Office LTSC 2021/2024) Leaks your password hash from a crafted message, already exploited Office update channel, not Windows Update. Then reset passwords
Windows domain controllers (LDAP, Active Directory) Remote code execution without logging in November 10 Windows security update
Exchange Server Remote code execution, and Windows Search indexing of attachments November Exchange update plus the Windows update. Test first: KB5131085 was linked to Exchange failures
.NET 8 / 9, PowerShell 7.4 / 7.5, Windows 11 23H2 No more security fixes after this month Move to .NET 10, PowerShell 7.6 and a newer Windows 11

Optimizely Web (website script)

There's no patch for you to install: Optimizely fixed its signup flaw itself on November 16. What protects your site now is cleaning up: remove code you didn't write from Project JavaScript, delete unknown users and API tokens, and rotate cookies. For visitors, the protection is an iPhone update: iOS 18.7.3+ or 26.3+.

Why "patch everything" can backfire

Here's the twist. Patches can break things too.

As of November 11, Qualys rated about 35% of November's patches as low reliability. One cumulative update, KB5131085, was reported to cause Exchange failures. Rushing every fix to every machine, untested, is how a security update turns into an outage.

Resilient Cyber calls this Patchmageddon: patching whatever AI finds, skipping the questions that matter. Is it actually being exploited? Can an attacker even reach it? Did we test it?

GreyNoise co-founder Andrew Morris adds a useful reality check: an exploit attempt is not a successful exploit. Fewer than 1 in 100 attempts against the big-name bugs actually work. Loud scanning doesn't mean you're breached.

And what about this month's huge outages? The Washington Sun reported, citing unnamed sources, that rushed AI-found patches caused some of them. No official record confirms that. Officially, the FAA blamed a new router, Microsoft a network defect dating back to July, and JPMorgan Chase a software bug. The Director of National Intelligence said it was not a cyberattack. We covered those in Microsoft, FAA, Banks: Why Everything Keeps Going Down.

A glowing mask in the dark. Most loud exploit attempts fail, but the few that work are the ones that matter.

A glowing mask in the dark. Most loud exploit attempts fail, but the few that work are the ones that matter. (Photo: sebastiaan stam, via Unsplash)

What to patch first: a simple order

You can't fix everything in one night, and you don't need to. Use this order, based on advice from Qualys, ManageEngine and Resilient Cyber:

  1. Known exploited first. Anything on CISA's KEV list. This month: the Outlook bug CVE-2026-124194. Aim for hours, not weeks.
  2. Internet-facing and identity systems next. Exchange, domain controllers, VPNs, remote access. Then check CISA's other urgent notes: Java installs, including Java bundled inside other products, and nginx (update it, or turn off HTTP/2 until you can).
  3. Easy, reliable patches on autopilot. Low-risk updates with a good track record can roll out automatically.
  4. Big or risky patches in rings. Test on a small group first, then widen.
  5. Can't patch yet? Mitigate. Turn off the vulnerable feature or isolate the machine, and remove the workaround once a fixed update ships.
  6. Delete what you don't use. Every unused app is one less thing to patch.

One more rule: "patched" means verified. Count a machine as done only when every required update is installed and a rescan confirms it.

What this means for you at home

You don't run a domain controller, but the same idea applies:

  • Turn on automatic updates for Windows, macOS, iOS, Android and your browser
  • Update Office or Microsoft 365 separately: open any Office app ‚Üí File ‚Üí Account ‚Üí Update Options ‚Üí Update Now
  • Restart when asked. Many fixes don't apply until you do
  • Be careful with "resend" or outage-themed emails. Scammers love a big news week
  • Use a password manager and two-step login, so a leaked hash or password isn't enough on its own

Late-night typing in a hood. At home, automatic updates and a restart do most of the work for you.

Late-night typing in a hood. At home, automatic updates and a restart do most of the work for you. (Photo: Muhamed Alaa, via Unsplash)

The bottom line

AI didn't invent new kinds of attacks. It made the old ones fast. That makes patching more important than ever, but "patch everything right now" is the wrong lesson. The right one is: patch what's exploited first, patch what's exposed next, test the rest, and check that it really worked.

A hooded figure on a neon-lit street. Attackers now automate the boring parts, so defenders have to automate theirs too.

A hooded figure on a neon-lit street. Attackers now automate the boring parts, so defenders have to automate theirs too. (Photo: Irham Setyaki, via Unsplash)

The next Patch Tuesday is December 8, 2026. Expect it to be big again.

Sources: FBI FLASH (via Resilient Cyber Brief #123), Qualys, ManageEngine, Mandiant M-Trends 2026, GreyNoise, Cloudflare, BTI, Calif, Xint, Keycloak, Orca Security, Cycode, StepSecurity, HeroDevs, Report URI. Patch counts differ by source because vendors count CVEs differently.

Originally published on rainytech.dev. This article was written with AI assistance and fact-checked against the sources listed above.

Read it in Khmer (ភាសាខ្មែរ): rainytech.dev/km

📰 Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.