Isometric cyber servers breached by glowing zero-day exploits.

Daily Cybersecurity News – November 20, 2026

CyberRecaps is supported by its readers. We may earn an affiliate commission at no extra cost to you if you buy through a link on this page.

Ghost in the Wires by Kevin Mitnick

Armory Castle exploits VMware vCenter zero-day since September

Critical

Affected product lines

  • VMware vCenter Server (standalone and Cloud Foundation / vSphere Foundation)
  • VMware SDDC Manager
  • VMware Tools for Windows
  • Telco Cloud Platform and Telco Cloud Infrastructure

What happened

A suspected China-nexus espionage group tracked as Armory Castle (formerly UNC5221) has exploited a VMware vCenter zero-day since at least early September, more than two months before the patch.

CVE-2026-123890 (CVSS 9.8) is an authentication bypass in the VMware Directory Service that lets a remote unauthenticated attacker obtain an administrator session on vCenter Server. Broadcom issued the fix on November 19 in VMSA-2026-0009. Google Threat Intelligence Group and Mandiant investigated multiple incidents; the group used a custom Go tool called BONESAW after initial access via a BeyondTrust Remote Support flaw, then SAWSNAKE to forge SAML tokens, bypass MFA, and move to ESXi hosts for credential theft.

Who is affected

Organizations running any version of VMware vCenter, SDDC Manager, or VMware Tools for Windows prior to the fixed releases, including deployments in VMware Cloud Foundation, vSphere Foundation, Telco Cloud Platform, and Telco Cloud Infrastructure.

Fixed versions include vCenter 9.1.1.0200, 9.0.2.0200, 8.0 U3m, and 5.2.5, plus corresponding Telco Cloud releases. Windows VMs with vulnerable Tools are also in scope.

Why it matters

vCenter is the central management plane for VMware virtualization estates. An unauthenticated path to admin sessions gives attackers full control over hosts and VMs, enabling mass credential theft, lateral movement, and persistent access that survives MFA.

Espionage groups chaining edge-appliance zero-days into core infrastructure shows how perimeter compromises rapidly become full environment takeovers. Operators must assume compromise if unpatched systems were internet-reachable or reachable from compromised appliances.

How it could have been prevented

Apply the Broadcom patches listed in VMSA-2026-0009 immediately for all affected vCenter, SDDC Manager, and Tools versions. There are no workarounds.

Hunt for indicators of BONESAW/SAWSNAKE activity, unexpected SSO administrator accounts, vCenter directory database copies, forged SAML tokens, and anomalous access from BeyondTrust or other edge appliances. Isolate management interfaces and review ESXi and AD database access logs.

Relevant professional terms

Authentication bypass
A flaw that lets an attacker skip the normal login process and gain access as if they were a legitimate user or admin.
SAML token forgery
Creating fake Security Assertion Markup Language authentication tokens by stealing the signing key, allowing impersonation of any user and MFA bypass.

phpAfterShell turns PHP session flaw into remote code execution

High

Fixed releases

  • PHP 8.2.36
  • PHP 8.3.37
  • PHP 8.4.28
  • PHP 8.5.13
  • PHP 8.6.0 (includes fix)

What happened

PHP's default file-based session storage can be forced to return attacker-chosen objects by interrupting a session write at the right moment. Combined with a suitable gadget chain in the application, this yields remote code execution.

The flaw, called phpAfterShell and tracked as CVE-2026-126714, received a 9.2 CVSS 4.0 score from PHP maintainers. Depthfirst discovered and reported it. Today's PHP releases fix it; the earlier October 21 releases that addressed PHP2Shell do not. No in-the-wild attacks have been reported.

Who is affected

Servers running PHP 8.2 before 8.2.36, 8.3 before 8.3.37, 8.4 before 8.4.28, or 8.5 before 8.5.13. PHP 8.6.0 includes the fix. End-of-life branches 8.1 and older are affected in unpatched builds but not in official EOL mitigation packages.

Only the default files session handler is vulnerable. Sessions stored in Redis, databases, or other backends that replace records atomically are outside this defect. Apps using upload progress or storing untrusted data in sessions are higher risk.

Why it matters

Default PHP session handling is extremely common. An interrupted write that leaves partial objects enables object injection leading to RCE when gadget chains exist, turning a storage race into full server compromise without authentication in many configurations.

Appliances and shared hosting that still ship older PHP or leave upload_progress enabled remain soft targets even after prior emergency patches.

How it could have been prevented

Install the fixed release for your PHP branch (8.2.36, 8.3.37, 8.4.28, 8.5.13, or 8.6.0) and confirm the running version. Rebuild and redeploy containers.

If immediate update is impossible and the app does not need upload progress, set session.upload_progress.enabled = 0 and restart workers. Limit untrusted data placed in sessions and consider switching to an atomic session store. The disable is only a partial mitigation.

Relevant professional terms

Session handler
The PHP component that saves and restores user session data between requests, most often by writing files on disk by default.
Object injection via partial write
A technique where an interrupted serialization leaves attacker-controlled object fragments in a session file that PHP later deserializes into live objects, enabling gadget-chain code execution.

Hackers breach Cisco Expressway devices hours after PoC release

High

What happened

Threat actors exploited CVE-2026-76466 in Cisco Expressway shortly after a public proof-of-concept appeared, compromising devices at a Russian organization and reaching its internal network.

The vulnerability (CVSS 9.8) stems from persistent cookies and insufficient client-certificate validation when TLS Verify Mode is enabled for clustering. An unauthenticated remote attacker with network access can connect to the database service, invoke RPC commands, and run arbitrary OS commands as root. Russian firm F6 observed the first known in-the-wild cases against X15.5.3 devices. Attackers took four devices (two internet-facing, two internal cluster nodes), then used custom Python scripts for SMB/SQL/WinRM scanning, GitLab access, and credential theft. Scripts show possible AI-assisted development. No attribution to a known group.

Who is affected

Cisco Expressway Series devices. Releases X14 and earlier require migration to a fixed release; X15.5.3 and earlier are fixed in X15.5.20 (patches released October 21).

Any organization with Expressway exposed or clustered with TLS Verify Mode enabled is at risk, especially those that delayed patching after the advisory and PoC release.

Why it matters

Expressway sits at the collaboration perimeter. Root RCE on it immediately yields a beachhead into internal clusters and adjacent services such as SMB shares, databases, WinRM, and source-control systems.

The speed from PoC to real breaches underlines that critical network appliances are hunted the moment exploit code is public. AI-assisted scripting further lowers the skill bar for post-exploitation.

How it could have been prevented

Upgrade to a fixed release (X15.5.20 or later supported version) at once. Audit for the published IoCs: IPs 66.179.31.32 and 18.130.47.116, "wombat@127.0.0.1" in erlang.log, and Python scripts under /mnt/harddisk/i{0-9}{4}[a-zA-Z]?/*.py.

Isolate compromised devices, scope lateral movement, rotate credentials, and restrict management and clustering interfaces. Block unnecessary outbound paths from Expressway.

Relevant professional terms

Proof-of-concept (PoC)
Public demonstration code that shows a vulnerability can be exploited, often accelerating real-world attacks once released.
Persistent cookie abuse
Leveraging long-lived authentication cookies that are not properly bound or validated, allowing an attacker to reuse them for privileged database or RPC access.

PHP2Shell hits firewalls as first exploitation target

Critical

What happened

Attackers quickly turned the PHP2Shell pair of vulnerabilities against internet-exposed Palo Alto Networks PAN-OS management interfaces. On October 28 Palo Alto confirmed exploitation of CVE-2026-109618 and CVE-2026-109619, both of which allow unauthenticated code execution on the web server.

Disclosure leaked early on October 20; PHP shipped an emergency patch the next day. The bugs are not trivial to exploit reliably on general servers because of ASLR, but appliances run identical binaries, letting attackers pre-compute offsets. Dream Lab (working with Palo Alto) found the actor obtained firmware, built memory profiles for nine versions, leaked libphp.so via Set-Cookie, and chained the flaws with Linux kernel CVE-2026-46300 (Fragnesia) to reach root on 25 systems (24 firewalls and one management server) across the U.S., Japan, Taiwan, and South Korea. Both PHP CVEs and the kernel flaw are in CISA KEV.

Who is affected

Organizations with PAN-OS management interfaces reachable from the internet, especially those still on end-of-life PHP versions shipped in the appliance (8.0.30 and 7.4.33 observed). Any network device or appliance embedding vulnerable PHP is in the broader blast radius.

General PHP web servers are also affected but harder to hit consistently when ASLR is effective.

Why it matters

Firewalls and management planes are high-value targets. Turning a web-stack bug into root on the device that is supposed to protect the network collapses the security boundary.

The episode illustrates that "everything is a web application," including critical appliances. Predictable firmware and outdated embedded runtimes make reliable exploitation far easier than on patched general-purpose servers.

How it could have been prevented

Ensure PAN-OS is fully patched and management interfaces are not exposed to the internet. Apply the PHP emergency fixes and any Palo Alto hotfixes. Monitor for the described memory-leak and cookie behaviors.

Treat all appliance web UIs as internet-facing attack surface: restrict access, keep embedded software current, and hunt for post-exploitation of kernel flaws such as Fragnesia. Disable or isolate unnecessary PHP features on appliances.

Relevant professional terms

ASLR
Address Space Layout Randomization - a defense that randomly arranges memory locations of code and data so attackers cannot reliably predict addresses needed for exploits.
Appliance firmware profiling
Extracting and analyzing the exact binaries and library layouts from device firmware so an exploit can be tuned to fixed offsets that are identical across every unit of that software version.

Gatekeeper asymmetry makes powerful LLMs more vulnerable

Medium

What to watch

  • Whether production LLM deployments upgrade input filters to match or exceed the power of the main model (including reasoning and tool use).
  • New bypass research similar to PuzzleMask that targets external AI safety layers.
  • Adoption of cryptographic or proof-based checks instead of pure heuristic gatekeepers.
  • How scaffolding and code-execution features change both attack and defense surface.

What happened

Check Point Research formalizes the "gatekeeper asymmetry doom" principle: if the system that vets adversarial input has strictly less power than the system that processes that input, and does not rely on a hard cryptographic or computational problem, compromise is inevitable.

The paper traces the pattern across 30 years of examples - hash-based antivirus, sandbox evasions, static classifiers, and their own PuzzleMask work bypassing external AI guardrails - and links it to results such as the Time Hierarchy Theorem. Frontier LLMs with scaffolding, extended reasoning, and code execution inherit the same risk while also offering a way to strengthen weaker gatekeepers.

Who is affected

Builders and operators of LLM applications, AI guardrails, content filters, sandboxes, and any security control that inspects input with a weaker model or simpler logic than the downstream system that acts on it.

Also relevant to classic AV, binary analysis, and web-application input validation teams.

Why it matters

Powerful models can be jailbroken or abused more easily when a weaker filter sits in front of them. The same asymmetry that defeated signature AV and many sandboxes now appears in AI safety layers.

Recognizing "doomed" surfaces early lets teams raise gatekeeper capability or narrow the power gap before attackers routinely bypass the checks.

Relevant professional terms

Guardrail
A filter or policy layer placed in front of an AI model to block harmful, off-policy, or adversarial prompts before they reach the main system.
Gatekeeper asymmetry doom
The principle that a weaker vetting system in front of a stronger processing system will eventually be bypassed unless the gatekeeper leverages a genuinely hard problem such as cryptography.

FBI poses as victim and attacker to bust ransomware scammer

Medium

What it means

Organizations hit by ransomware should treat unsolicited recovery offers with extreme caution, verify any service through trusted channels, and report suspicious approaches to law enforcement. The case shows authorities are actively policing the secondary market that springs up around ransomware incidents.

What happened

Federal prosecutors say the FBI ran an undercover operation in which agents posed as both a ransomware victim and the threat actors themselves to catch an alleged fraud scheme.

The target was a Florida-based company run by Juan Ricardo Palacio Escobar and an unnamed associate. The firm, started in 2020, claimed to offer data-recovery services for ransomware victims. By controlling both sides of the interaction the FBI could observe the scheme in real time and build the case.

Who is affected

Ransomware victims who might have been solicited by or paid the Florida recovery firm, and the broader market of third-party "decryptor" or recovery services that operate in the gray zone after attacks.

The indictment focuses on the operators of that specific company.

Why it matters

Fake or abusive recovery services prey on organizations already in crisis, often collecting fees while delivering little or simply recycling attacker access. Law-enforcement ability to insert itself as both victim and operator demonstrates a powerful tactic against this secondary fraud layer.

It also signals increased scrutiny of the post-ransomware ecosystem, not only the initial extortion groups.

Relevant professional terms

Ransomware recovery service
A third-party company that offers to help restore data or negotiate after a ransomware attack, sometimes legitimately and sometimes fraudulently.
Undercover dual-role operation
A law-enforcement technique in which agents simultaneously impersonate the victim and the criminal side of a transaction to capture the full fraudulent scheme as it happens.
Source: 404 Media

European Commission confirms AI tool used in cyberattack

High

What happened

The European Commission confirmed it suffered a cyberattack detected on October 22 that involved a publicly available AI tool. Spokesperson Thomas Regnier called it an "AI-generated cyber campaign."

POLITICO reported that CERT-EU identified use of ARTEX, an open-source AI hacking platform. The FBI had warned the previous week that ARTEX is employed by both criminals and state-sponsored groups to automate vulnerability discovery and exploitation, sometimes reaching privileged access in under three hours. The attack disrupted personnel systems (payroll, leave, medical reimbursements, translation) for more than 10 days. Several services were taken offline as a protective measure. No broad staff data breach or ransom demand has been found; the investigation continues. This is the Commission's third disclosed incident in 2026.

Who is affected

European Commission staff and services that rely on the affected personnel, payroll, leave, medical, and translation systems. Potentially other EU bodies monitored by CERT-EU.

Broader implications for any organization whose internet-facing systems can be probed by AI-driven exploitation frameworks such as ARTEX.

Why it matters

Confirmation that a major governmental institution was hit with an AI-orchestrated campaign marks a practical escalation from theoretical warnings. Open-source tools that wrap commercial models lower the skill and time required to find and exploit flaws.

Disruption of core HR and administrative systems for over a week shows operational impact even without confirmed mass data theft. Defenders must assume AI-assisted reconnaissance and exploitation are now routine.

How it could have been prevented

Patch and harden internet-facing systems aggressively; ARTEX-style tools excel at rapidly identifying known and weak vulnerabilities. Segment personnel and administrative networks, enforce least privilege, and monitor for automated scanning and exploitation patterns.

Review logs around October 22 for anomalous access, take vulnerable services offline when under active attack, coordinate with CERT-EU or equivalent, and notify affected individuals and data-protection authorities promptly.

Relevant professional terms

CERT
Computer Emergency Response Team - an organization that coordinates the response to security incidents, such as CERT-EU for European Union institutions.
AI-driven exploitation framework
An open-source or commercial platform that uses large language models to automate vulnerability discovery, exploit generation, and privilege escalation at machine speed.
Source: The Record