Since 21 October, PHP, OpenSSH on Alpine, Apache httpd, the Apache Axis library, Java and nginx have all had critical flaws, and for four of them attacks were reported within days. Several were found with AI, in hours rather than months. On 17 November Google described the other side: an AI agent that reads security patches, writes the exploit and runs it, with no human in between. The patch itself has become the starting signal. Anyone who does not know by that evening where nginx, Java and Apache run in their own company will lose this race.
Four weeks in November
We wrote about PHP2Shell on 21 October and about SSHole in OpenSSH on 6 November. Both times attackers were faster than most administrators: PHP2Shell was attacked on the morning of the release, SSHole less than a day after the fix. It did not stop there. In the two weeks since, Apache httpd, Apache Axis, Java and nginx have followed, together with a handful of other critical flaws that the US agency CISA lists as actively exploited.
21 October to 18 November 2026
Four weeks, six critical flaws in software that runs everywhere
- 21 OctPHP2ShellTwo flaws in PHP itself. Details circulated hours before the release, attacks the same morning.
- 5 NovSSHole in OpenSSH on AlpineA 257-byte password opens accounts that have none. Root on a honeypot less than a day later.
- 6 NovApache Axis 1.xUnauthenticated code execution in a library that is end of life and sits inside Confluence and N-central.
- 11 NovApache httpd 2.4.71Fixes heap corruption in mod_http2, versions 2.4.55 to 2.4.70.
- 13 NovAxis exploited in ConfluenceRapid7 sees hundreds of encoded commands, probably from AI-generated scripts, and a Microsoft 365 sign-in after 12 hours.
- 17 NovGoogle: an AI agent attacks on its ownexploit_agent compares code before and after patches, writes exploits and runs them, with no human in the loop.
- 17 NovJava4ShellA bug in the JVM's Base64 decoder, found with Codex in about an hour, working exploit two hours later.
- 17 Novhttpd under attackSingapore's Cyber Security Agency warns of active exploitation and a public proof of concept.
- 18 NovNGINX2RootFive flaws in nginx, four of them found by an automated scan, plus a Linux kernel bug that leads to root.
- 18 NovBACS and CISAThe Swiss federal office warns about Java4Shell. CISA lists Axis as exploited, Puppet Enterprise a day earlier.
What these flaws have in common is not the type of bug but where they sit: in web servers, runtimes and libraries that are part of almost every product, often without anyone having chosen them. Very few companies have Axis on a list. Many have Confluence.
NGINX2Root: from visitor to root
On 18 November nginx released versions 1.31.9 and 1.30.7, which fix five flaws that the security firm Xint calls NGINX2Root. The widest one, CVE-2026-123007, is a memory corruption in HTTP/2 flow control. According to Xint's write-up, the code is in every release from 1.9.5 to 1.31.8, and all it needs is HTTP/2 on a TLS listener, which is what most setups use today. Without a login, an attacker runs code as the nginx worker. The other four affect the resolver, HTTP/2 connection draining, HTTP/3 and the mail proxy, each only with a particular configuration.
The step to root is not nginx's doing. Xint chains a bug in the SCTP code of the Linux kernel, present since 2018, and reports a root shell on the host even with nginx in a Docker container. There is no kernel fix yet. Until there is, the module can be blocked with install sctp /bin/false in a file under /etc/modprobe.d/. That does not help if SCTP is already loaded or built into the kernel.
Xint found the first flaw by hand while preparing for a hacking competition. The other four came a few days later from a fully automated scan of its AI-based platform, without human guidance, Xint says. From discovery to release took 38 days. No attacks have been reported so far, and Xint is holding back the final steps of the exploit.
Two details matter in practice. Distributions backport fixes without changing the version that nginx -v reports, so check the package changelog. And a reload is not enough: old workers keep their connections, so nginx has to be restarted. nginx also runs in places nobody thinks of as nginx: Kubernetes ingress controllers, appliances, NAS devices and the reverse proxy in front of almost every web app.
Java4Shell: the flaw is in the runtime
A day earlier, Oracle and the OpenJDK project fixed CVE-2026-114938, which its finders at Calif call Java4Shell. The bug is in the hand-written assembly that speeds up Base64 decoding on x86-64 processors with AVX-512 VBMI, which includes most current Intel Xeon and AMD EPYC servers. A crafted Base64 string writes past the end of a buffer. Web applications decode Base64 all the time: in login headers, tokens and uploads. That is how Calif showed code execution without a login in Keycloak, Jira, Confluence, JFrog Artifactory, Jenkins and PingFederate, and with a login in Kafka, Solr, Elasticsearch and others.
Affected are all OpenJDK-based builds from 17.0.6 on, including 21, 25 and 27, whether from Oracle, Temurin, Corretto, Microsoft, Azul, SAP or Red Hat. The fixed versions are 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1. According to Qualys, the flag -XX:-UseBASE64Intrinsics switches to the regular decoder in the meantime. Both need a restart of the Java process. On 18 November the Federal Office for Cyber Security (BACS) asked operators to update immediately and to check for compromise, because an update does not remove an attacker who is already inside.
How the flaw was found is the real news. According to Calif, Quang Luong found it with OpenAI's Codex in about an hour and had a first working exploit two hours later, for under 200 dollars in tokens. No attacks were known on 17 November. Qualys expects others to repeat the work, and reports that, according to the researchers, details leaked before the patch.
Apache, Confluence and Puppet: already under attack
For three other flaws the question is no longer whether attacks will come:
- Apache httpd, CVE-2026-119618. A race condition in mod_http2 corrupts the heap in versions 2.4.55 to 2.4.70. Apache fixed it in 2.4.71 on 11 November. Two teams had reported it independently, one through the open-source scanner of the AI company Anthropic, the other from Google. On 17 November Singapore's Cyber Security Agency warned of active exploitation and a public proof of concept. Servers without HTTP/2 are not exposed, so
Protocols http/1.1is the stopgap. - Apache Axis 1.x, CVE-2026-90436. The library is end of life, and Apache will not release a fix for 1.x. It still sits in Confluence Data Center, if the remote API is switched on, and in N-able N-central. Rapid7 saw the flaw exploited against Confluence on 13 November: hundreds of hex-encoded shell commands, structured in a way that suggests AI-generated scripts, and after about 12 hours a sign-in to Microsoft 365 with a service account. CISA added it to its list of exploited flaws on 18 November. Fixed in Confluence 9.2.28 or 10.x and N-central 2026.4 HF3.
- Puppet Enterprise, CVE-2026-115731. Agents with a signed certificate can run code on the Puppet server, and if autosigning is on, possibly anyone. CISA has listed it as exploited since 17 November. A Puppet server controls the configuration of every machine it manages.
Since 18 November the same CISA list also contains CVE-2026-40217 in LiteLLM, a widely used gateway for AI models. According to Google, attackers use it to turn limited API keys into admin access and run up model usage at the victim's expense. AI infrastructure is now a target in its own right.
AI on both sides
On 17 November Google's Threat Intelligence Group published a report that changes the timing for everyone. On an exposed staging server, Google found a framework called exploit_agent, run by a suspected Russian-speaking operator. It compares popular web software before and after a patch, works out what was fixed, writes an exploit and runs it without human intervention. The server held data already stolen from victims, mostly French websites in media, entertainment and education. Google says it is the first time it has seen discovery and exploitation automated without a human in the loop.
exploit_agent, after Google's report
From patch to break-in, with no human in between
- A vendor publishes a patchThe difference between the old and the new version shows where the flaw was. For a model, that diff is a map.What stops it: An inventory that already exists, so you know within minutes whether you are affected.
- The agent compares the codeIt watches popular web software continuously and analyses every change for a fixed vulnerability.
- It writes the exploitFor Java4Shell, the finders needed about three hours and under 200 dollars of tokens. Attackers do not need to be faster than that.What stops it: A patch deadline in hours for exposed systems, and an emergency change that does not wait for the next window.
- It runs the exploit against targetsAutomatically, against everything it can reach. Google's victims were French media, entertainment and education websites.What stops it: No admin console, build server or login service openly on the internet.
- Webshell, brute force, data theftThe staging server held scripts for reconnaissance and webshells, and data already stolen from victims.What stops it: Alerts when a web server or Java process starts a shell, new files in the web root, unusual outbound connections.
It is not an isolated case. The report names a series of agent frameworks that attackers use for multi-stage intrusions, among them ARTEX, CAIRN, CyberStrikeAI, HexStrike AI and PentAGI. Zero-day exploits rose from an average of 8 a month in 2025 to 11 in 2026, and to 22 in August alone. Even before that, according to VulnCheck, almost 29 percent of the flaws known to be exploited in 2025 were attacked on or before the day their CVE was published. Google's conclusion for defenders is unspectacular: most intrusions still go through stolen credentials and unpatched systems, and the agents trade stealth for speed, which leaves traces that can be detected.
The same tools work for defenders. Java4Shell, four of the five nginx flaws and one of the two reports for httpd came from AI-assisted or automated searches by researchers who reported them responsibly. Use them on your own code and configuration before someone else does. But treat them as attack surface too: according to Google, a suspected North Korean group changed the settings of the coding assistant Claude Code on developers' machines so that it downloaded a backdoor.
You can only patch what you know about
In practice, a Java4Shell evening goes like this: the warning comes in, and someone has to find out where Java 17 or newer runs on an AVX-512 machine. On the build server, sure. But also in the Keycloak container, in the Confluence installation, in the Elasticsearch behind the website search, and in the appliance whose vendor bundles its own JDK without saying so. The BACS and CISA both explicitly mention components inside containers and appliances, which are missing from most inventories.
The flaws of the last four weeks show what an inventory needs to help on the evening it is needed:
- Everything that faces the internet, with host name, owner and what runs on it. The attackers' agents scan this first, and it has to be patched in hours, not weeks.
- Runtimes and libraries, not just products. Java, PHP, nginx, OpenSSH, Axis. A software bill of materials (SBOM) from your own build or from the vendor shows what is inside.
- The real version. The package version and its changelog, not what the program reports about itself, because distributions backport fixes.
- Who is responsible and who decides. An emergency patch at 10 pm needs a name and the authority to restart a service.
Without such a list, every new warning starts from scratch. The machine on the other side does not.
Security is a management job
None of this is a hobby for the IT team. In Switzerland the board of directors is responsible for the overall management and organisation of a company (Art. 716a CO), and today that includes knowing which systems face the internet and how quickly they get patched. The Data Protection Act requires appropriate data security (Art. 8 FADP), and anyone who intentionally ignores the minimum requirements of the ordinance can be fined up to 250,000 francs personally (Art. 61 FADP). Since 1 April 2025, operators of critical infrastructure have had to report cyberattacks to the BACS within 24 hours. And anyone who sells products with software in the EU, from Switzerland too, has had to report actively exploited flaws in those products since 11 September 2026 under the Cyber Resilience Act: an early warning within 24 hours, a notification within 72.
Responsibility does not end at your own servers either. The JDK in a vendor's appliance is the vendor's to fix, but the risk is yours. Ask your suppliers whether their products contain nginx, Java 17 or newer, or Axis, and when the fix will come, and write patch deadlines into your contracts.
What to do now
What to check this week
Fixed versions and stopgaps
- NGINX2Root, 5 CVEsnginx
- Fixed in 1.31.9 and 1.30.7 of 18 November
- Until then: HTTP/2 and HTTP/3 off, no resolver with request variables
- Restart, a reload keeps old workers running
- Java4Shell, CVE-2026-114938Java 17.0.6 and later
- Only on x86-64 CPUs with AVX-512 VBMI
- Fixed in 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1
- Until then: -XX:-UseBASE64Intrinsics, then restart
- CVE-2026-119618Apache httpd with HTTP/2
- mod_http2 in 2.4.55 to 2.4.70
- Fixed in 2.4.71 of 11 November
- Until then: Protocols http/1.1 in every virtual host
- Apache Axis, CVE-2026-90436Confluence Data Center and N-central
- Confluence: 9.2.28 or 10.x, otherwise turn off the remote API
- N-central: 2026.4 HF3, which disables the Legacy API
- Look for compromise from 12 November
- From October and early NovemberPHP and OpenSSH on Alpine
- PHP2Shell: PHP 8.5.12, 8.4.27, 8.3.36 or 8.2.35
- SSHole: OpenSSH 10.6p2 or Alpine's fixed package
- Restart PHP workers and sshd
- No fix yetLinux kernel, SCTP
- The root step of NGINX2Root, in the kernel since 2018
- Block the module: install sctp /bin/false in /etc/modprobe.d/
- Does not help if SCTP is already loaded or built in
- Patch exposed systems first, today. nginx, Java, Apache httpd, Confluence, N-central and Puppet, starting with what faces the internet and with login, build and collaboration services. Restart the processes: a fixed package does not reach a running JVM or old nginx workers.
In CyberTell: the Vulnerability radar (beta) shows which entries in your inventory are affected by a CVE, with the fixed version and a button for a ticket. Flaws on CISA's list of exploited vulnerabilities are always marked critical. - Look for compromise, not just for the version. For Axis from 12 November, for httpd from 11 November: web server or Java processes that start shells, new files in the web root, unusual outbound connections and sign-ins from service accounts. If you find something, treat it as an incident, not as a patch.
In CyberTell: an Incident response case keeps evidence and the chain of custody together, and the War Room gives the team a crisis channel that does not depend on your own mail. - Build the inventory you would have needed tonight. Internet-facing systems, runtimes and libraries, real package versions, owners. Import SBOMs wherever you have them.
In CyberTell: the Vulnerability radar accepts CycloneDX and SPDX files, lock files and Dockerfiles, and checks new CVEs against them every night. Device Management reports the installed software of each computer, and the radar suggests systems from it. - Know your attack surface from the outside. Which hosts, ports and admin consoles are visible from the internet, and which ones nobody remembers.
In CyberTell: Attack surface (beta) maps it passively from public sources once a day and reports new subdomains, risky ports, admin panels and known CVEs on exposed services. - Set patch deadlines and keep them. Our suggestion: exposed systems with an exploited flaw within 48 hours, other critical flaws within 14 days, and an emergency change that does not wait for the next maintenance window.
In CyberTell: a Playbook records the procedure for an emergency patch with roles and steps, and a Tabletop exercise lets the team practise the evening. - Take admin consoles off the internet. Jenkins, Jira, Confluence, Artifactory, Keycloak and the admin interfaces of appliances belong behind a VPN or an access proxy.
- Ask your suppliers. Whether their products contain nginx, Java from 17.0.6 or Axis 1.x, and by when they will deliver the fix.
In CyberTell: under Suppliers you send a compliance request that the supplier confirms, and keep their answers and SLAs on file. - Prepare people for AI-made attacks. Phishing mails without spelling mistakes, voice messages in the boss's voice. The technique is new, the defence is not: a second channel to check before anyone pays or hands over a password.
In CyberTell: Phishing campaigns personalise templates with AI, Vishing campaigns send voice messages with a cloned voice via WhatsApp or SMS, and the course "Recognising Deepfakes & AI Fraud" shows what to look out for.
The notes "In CyberTell" refer to CyberTell, our own platform. Every step works without it.
Do you know what your company exposes to the internet? Pentesting and consulting →