Three stories set the tone for the week. An intermediate certificate authority turned out to have private keys so weak that one of them was the number 4. A shared AI inference service returned one customer’s text inside another customer’s answer. VMware’s vCenter had a 9.8 authentication bypass that a suspected state-linked group had been using since early September, more than two months before a patch existed.
Reactions sorted themselves into three camps. Criminal gangs, said the first. AI, said the second. Foreign governments, said the third.
All three explanations have evidence behind them this week. Each also answers a narrower question than the one people are asking, and only one of those answers produces work a defender can finish before the next incident.
TL;DR
- The week’s incidents were technically unrelated: an exploited vCenter zero-day, memory corruption in OpenJDK’s Base64 decoder, a skimmer injected into a widely used marketing script, five NGINX flaws, a wiped ESXi estate, a weak-key CA compromise, a cross-tenant KV-cache leak, and agentic reconnaissance frameworks.
- N-central is the week’s correction to a hopeful story. Researcher-disclosed 4EverAble2Shell flaws were used to plant miners on 21 servers, and 19 of them were already running Hotfix 2.
- The strongest attribution claim is a vendor or government assessment. GTIG and Mandiant link the vCenter exploitation to ARMORY CASTLE, which they describe as PRC-nexus. The FBI attributes FaceStealer to North Korean actors it tracks as BlueNoroff. Check Point says it has not established who attacked its gateways.
- AI appears in three roles: offensive frameworks that compress reconnaissance, assistants that find bugs faster than vendors patch them, and a failure surface inside inference infrastructure.
- Project Zero cut its disclosure grace period on the argument that AI-assisted discovery and exploitation made it obsolete. The practical effect is that patch latency now costs more than it did last year.
- The recurring conditions were exposed management interfaces, concentrated trust in signing keys, scripts and certificate authorities, slow patch uptake, and recovery systems that depend on the systems they are meant to restore.
- Plan the week around what an intruder can reach and what they inherit, not around whichever label the headlines chose.
What happened this week
vCenter, exploited before it was patched. Broadcom published VMSA-2026-0009 on November 19 with a fix for CVE-2026-123890, a 9.8 authentication bypass. GTIG and Mandiant tied exploitation in early September to ARMORY CASTLE, formerly tracked as UNC5221. The toolset abused the vmdir DCE/RPC interface on TCP 2012, recovered an SRP session key through a predictable UUIDv1 timestamp, pulled the directory database, and used the stolen SAML signing key to forge assertions as an administrator. That is a compromise of the identity provider for the virtualization platform, and patching does not reset the trust it created.
N-central: the hotfix was not the boundary. N-able shipped Hotfix 2 on November 10, and the following week showed what it did not cover. Researcher Sean Bales disclosed 4EverAble2Shell, an unauthenticated chain that pairs CVE-2026-90436 (unsafe class instantiation in the bundled Apache Axis SOAP fault path, rated 9.8) with an N-central privilege escalation to root. Huntback found XMRig miners on 21 N-central servers in seven countries. Nineteen of the 21 were already running Hotfix 2, and the operators reached them through the Legacy API. N-able disabled that API by default and began a phased Hotfix 3 rollout on November 16.
A skimmer delivered from a trusted CDN. On November 16, an attacker wrote malicious JavaScript into Optimizely Web Experimentation customer projects, and Optimizely’s CDN served it for up to 17 hours. The incident, tracked as OPTIMJACK, reached 178 organisations, or 188 infected snippets out of 1,611 checked. The payload included a card skimmer, a crypto address clipper, cookie theft and iOS exploit kits. No customer server was breached. The code arrived from a domain every affected site had chosen to trust, with a valid certificate, past content security policy allowlists and past whatever reputation check the vendor had. Card fields inside payment-provider iframes stayed out of reach; card fields on the merchant’s own page did not.
Five NGINX flaws in a single advisory. F5 released NGINX 1.31.9 and 1.30.7 on November 18 for five vulnerabilities that allow unauthenticated code execution as the worker process. The headline bug, CVE-2026-123007, is an integer overflow in HTTP/2 flow control that becomes an out-of-bounds heap write; the rest are use-after-free flaws in connection draining, QUIC streams, the DNS resolver and the mail proxy. CERT-EU rates each 8.1 under CVSS v3.1. Wiz, one of five teams that found the bugs, demonstrated code execution on CVE-2026-123007, which dates back to NGINX 1.9.5 in 2015, and says its AI tooling found both HTTP/2 flaws. No exploitation in the wild has been reported. Disabling HTTP/2 mitigates two of the five.
Ten days to wipe a virtualization estate. Kaspersky published a case on November 13 in which an attacker destroyed four ESXi hosts, their virtual machines, two vCenter appliances and the backup infrastructure. No hypervisor flaw was involved. Credentials recovered outside the virtualization layer reached vCenter, and a root SSH login was followed within 43 seconds by the installation of MeshAgent, a legitimate remote management agent. The attacker then used vCenter Host Profiles to enable SSH on the hosts. Before the final stage, antivirus in the backup environment was disabled and recovery systems were encrypted. Ransom notes named BlackNevas, though Kaspersky did not recover the binary, so the family match rests partly on command-line style.
FaceStealer and the $400 million. The FBI published an alert and the US government seized cloud infrastructure behind FaceStealer, a North Korean campaign that ran for 18 months, stole more than $400 million in cryptocurrency and used over 900 secretly recorded webcam videos. The FBI attributes the activity to the group it tracks as BlueNoroff, also called Sapphire Sleet, UNC1069 or APT38. Operators hijacked accounts on Telegram, LinkedIn, Slack and Discord and invited targets to video meetings, where they played recorded footage of someone the target knew, blamed a technical problem and offered a link to a “fix”. The FBI says the group has shifted toward open-source maintainers, aiming at the software supply chain.
Disclosure windows got shorter. On November 13, Google’s Project Zero replaced its “90 + 30” policy with a “Full 90” one, dropping to 60 days in February 2027 and 30 days in January 2028. The 30-day soak period after a fix is gone, along with most special cases. Project Zero’s stated reason is that LLMs have increased both the speed of vulnerability discovery and the speed of exploiting a patch, to the point where “zero patch-to-exploitation time is a reasonable, though imperfect, model of attacker behavior”.
Weak CA keys at eMudhra. Eight intermediate CA private keys were recoverable from their public keys, one of them the integer 4. Mozilla blocked the certificates and shipped additional updates for clients that might miss the blocklist. Cloudflare saw roughly 160 handshakes presenting apparently forged certificates, assessed them as default server certificates, and reported no evidence of interception of its own traffic.
A cross-tenant AI leak. Cloudflare disclosed that its GLM-5.3 deployment returned 155 responses containing other customers’ data between August 27 and September 9. The advertised context window of 1,048,576 tokens exceeded the prefill KV-cache pool of 681,280 tokens, and an SGLang error path let decode continue with stale pages. A configuration mismatch, not an adversarial prompt.
Two incidents with no attacker in the room. A bug from a July update to Microsoft’s 365 gateway servers, plus a Teams failover and about 14% of capacity already offline, took Exchange Online mail down for roughly a day on November 12 and 13. Microsoft says its load tests missed the defect because they ran on hardware with fewer cores, and restarting servers did not help at first because the returning traffic overwhelmed them. Phishing posing as service-health notices followed within hours, citing the real incident ID. The same week, researchers Shubham Shah and Sam Curry found an OpenIDM console behind two carmakers’ ForgeRock login systems still using the documented default of admin and admin, and used it to start the engine on consenting owners’ cars.
Three explanations, three questions
The blame divides neatly because each explanation is true at its own layer.
The actor explanation answers who acted and what they wanted. It is the layer intelligence services and vendors argue about, and the layer with the weakest public evidence in most weeks. This week the vCenter attribution rests on a vendor assessment that later reporting may revise, the N-central exploitation has no named actor at all, Check Point declined to name one, and the FBI put a specific name and a dollar figure on FaceStealer. That range is normal. It sets the confidence level a defender should carry into planning.
The AI explanation answers how fast and how much. Automated frameworks compress the reconnaissance and privilege-probing phases that used to occupy an operator for days. Project Zero’s policy change is a vendor-independent acknowledgment that the same acceleration applies to finding bugs and to reading patches. A defender who sees 27 permission checks in 23 seconds has a correlation opportunity, not a signature.
The structural explanation answers why it worked. A forged SAML assertion works because vCenter is the trust anchor for the platform around it. A weak CA key works because clients inherit trust from stores nobody inventories. A stale KV page reaches the wrong customer because a capacity number was configured in two places. A skimmer reaches 178 sites because a marketing script is allowed to execute with the same authority as the page that embeds it.
The structural layer is the one that produces tasks: a signing key to rotate, a port to close, a cache pool to size, a decoder to replace.
Where the AI story holds
AI appeared in three separate roles this week, and merging them hides all three.
As tooling, it showed up in GTIG’s reporting on agent-driven frameworks in live intrusions, including CAIRN and a modified PentAGI-based platform. The frameworks ran reconnaissance, tested permissions, retried failed routes and adjusted after errors. What they leave behind is a log pattern: one valid identity touching many resources in a short window, then changing something that persists.
As a bug-finding accelerant, it shows up in Wiz’s account of the NGINX flaws and in Project Zero’s reasoning. Project Zero says it now finds several vulnerabilities a week in targets where it once found a few a year, and that the volume has made vendor-by-vendor grace periods unmanageable. Wiz says AI tooling found both HTTP/2 bugs. Neither claim says an AI can exploit a patched target on demand. Both say the interval between discovery, disclosure and exploitation is being compressed from the research side, not only the attacker side.
As a failure surface, AI systems produced ordinary infrastructure bugs with an unusual blast radius. The SGLang leak was a capacity check and a page-ownership error in shared memory, the sort of defect that has shadowed multi-tenant systems for decades. It reached production because the advertised context length and the actual cache pool were maintained separately, and because the rejection path was not tested at the boundary. No prompt injection was required. Describing it as an AI attack hides what needs fixing, because inference infrastructure now carries the same tenant-isolation obligations as any other shared service.
Nothing in this week’s reporting suggests AI has replaced exploit development. The vulnerabilities in these incidents were ordinary entries on a patch list, and the exploits that mattered still needed a reachable service and either a valid credential or a memory bug reliable enough to use.
Where the foreign-actor story holds
The vCenter case shows the state-nexus model at its most concrete: patient access, a zero-day held in reserve, and a willingness to be discovered months later. ARMORY CASTLE’s earlier exploitation of BeyondTrust Remote Support, followed by a pivot into vCenter, is the kind of chained campaign that commercial criminals rarely sustain.
FaceStealer is the better-documented example, and the more instructive one. The FBI did not say “foreign actors” and leave it there. It named the group, counted the money, described the technique and published indicators. The campaign also shows where state interest and supply-chain risk meet: the same infrastructure used to drain crypto wallets is now aimed at the maintainers of libraries that millions of systems install.
But “foreign” is a category, not a control. It says nothing about which of your interfaces is reachable, and it invites organizations to file an incident under geopolitics and stop asking engineering questions. It also encourages a common analytical error: fusing unrelated events into one campaign because the labels rhyme. This week’s incidents share conditions, not a command structure. Check Point’s refusal to attribute is also worth reading as a finding: an unanswered attribution question stays unanswered rather than getting filled in by inference.
The conditions behind the week’s incidents
Five conditions recur across these incidents.
Exposure. N-central, Check Point’s portal, vCenter’s SSO interface, the NGINX workers behind a dozen appliances, and the AI inference pool all control or serve other systems. In each case the party that caused the damage could already reach the control surface, through a path the operator believed was narrow.
Trust. It hides in a handful of keys and a handful of scripts: a SAML signing key, an intermediate CA key, an RMM agent certificate, a service account with replication rights, a single line of vendor-supplied JavaScript on a checkout page. Whoever holds one of those gets authenticated by everything downstream, and the failure looks like normal traffic.
Patch latency. N-able shipped Hotfix 2 and 19 of the 21 known victims were running it. NGINX’s worst flaw in this batch was introduced in 2015, inside a component many operators do not know they run: ingress controllers, appliances, hosting stacks. The gap between the advisory and the running process is where the incidents live, and Project Zero has just made that gap more expensive.
Recovery is a dependency, not a backup. The BlackNevas case destroyed backups first, then production. “We have backups” describes a path that includes the accounts, the management plane and the network needed to use them, and every one of those is a target.
Volume. The GTA VI campaign needed no zero-day and no state sponsor, only a release date and a search engine. The automakers’ default admin password needed no exploit development, only the knowledge that the console exists.
What You Can Do Today
- Inventory trust anchors, not just assets. List the SAML signing keys, intermediate CA certificates, RMM agents, backup service accounts and CI/CD tokens that other systems accept without further proof. Assign an owner and a rotation procedure to each, and test that rotation in a non-production environment.
- Treat vendor JavaScript as executable code on your checkout page. Inventory every third-party script and its version, compare the served file against a pinned hash where the vendor supports it, and check whether payment fields sit on your page or inside a provider iframe. An allowlist that names a trusted CDN will not stop an attacker who compromises that CDN.
- Verify patches against running processes and retired features. Confirm the fixed build is what the service loaded, including containers, bundled runtimes and standby appliances, and check that the interfaces an attacker used are actually disabled. Hotfix 2 did not close the Legacy API path, and a patched N-central was still reachable.
- Test recovery as its own attack path. Give the backup platform its own credentials and management network that production cannot reach in either direction, and prove a restore works with those credentials alone. Copy ESXi logs off the host, because they live in a ramdisk that a reboot clears.
- Hunt bursts by identity. Correlate rapid permission probes, repeated errors, test writes and deletes, and endpoint enumeration with the same principal’s later successful actions. Baseline deployment automation so the alert queue stays reviewable.
- Assume shorter disclosure windows in your patch plan. For each internet-facing product, decide in advance who can approve a workaround when no patch exists, and how a configuration change reaches production within days rather than quarters.
- Split the post-incident conversation. Record separately who the actor was, what tooling accelerated them, and which control failed. Only the third item becomes a change with an owner and a date.
Related Posts
- State-Sponsored Threat Actors 2026: Who They Are and What They Do
- FortiBleed’s Ransomware Pipeline: What SOCRadar’s INC Ransom and Lynx Attribution Actually Proves
- N-days Are Becoming N-hours
- Why Enterprise VPN and Gateway Products Are Perpetually Broken
- Week 37 Security Priorities: Revoke Trust Before You Chase Malware
Sources
- Broadcom — VMSA-2026-0009
- Google Threat Intelligence Group and Mandiant — ARMORY CASTLE Exploits vSphere
- Google Threat Intelligence Group — Agentic Threat Activity Crosses a Significant Threshold
- The Hacker News — Hackers Exploit N-central Flaws to Infect 21 Servers, Including 19 Running Hotfix 2
- ScruteX — Optimizely Supply-Chain Attack (OPTIMJACK): Full Analysis, IOCs and Victim List
- CERT-EU — Security Advisory 2026-022: Critical vulnerabilities in Nginx
- F5 — K000164058: NGINX security advisory
- Project Zero — Policy and Disclosure: 2026 Edition
- FBI — FaceStealer Cryptocurrency Theft Scheme
- BleepingComputer — Microsoft says months-old bug caused global Microsoft 365 outage
- DIESEC — Top 5 Cybersecurity News, November 20, 2026
- Calif — Java4Shell research and technical write-up
- OpenJDK — Security Advisory, November 17, 2026
- Check Point — Security Advisory: Active Exploitation of PHP2Shell and Local Privilege Escalation
- N-able — Update on N-central 2026.4 HF2 and ongoing security response
- Cloudflare — Cross-request KV cache exposure in the GLM-5.3 Workers AI model
- Mozilla Security Blog — Protecting Firefox users following the compromise of eMudhra intermediate certificates
- Cloudflare — Tracking exploitation and patching of eMudhra’s weak keys
Useful read?
Find us again on Google.
Add Hive Security as a preferred source for practical security research and analysis.
Add as a preferred source on GoogleChoose Hive Security in Google's source preferences.