9.2

CVSS score of PHP2Shell, the way in

7.8

CVSS score of CVE-2026-123039, the way to root

7

days between the hotfix and the first attack

2

attack waves, on 30 Oct and 2 Nov

The firewall runs a web server, and it needs patching like one

A Check Point Security Gateway with the Mobile Access portal switched on is a firewall with a public web application bolted to its front. That application runs on PHP. When PHP shipped a fix on 21 Oct 2026 for two flaws that allow code execution without a login, the gateway inherited the bug, and the patch had to come from Check Point, not from php.net.

It did come. Check Point says fixes were available on 23 Oct. Attackers started on 30 Oct anyway, against gateways that had not taken the hotfix, and by 4 Nov a couple of dozen systems were compromised. In a limited number of cases the intrusion did not stop at the firewall. It reached a VMware ESXi host and ran a wiper.

Our argument is simple. The firewall is not a sealed box with a version number. It is a Linux server running a web runtime, a stack of daemons and a set of credentials that reach the rest of the estate, and it needs the same patch speed and the same hunting that any internet facing server gets. The teams hit here were not undone by an exotic bug. They were undone by a seven day gap and a device nobody was watching from the inside.

We wrote up the PHP flaws themselves when they landed, in our analysis of the PHP core RCE. That post ended on the copies of PHP that sit inside products and never show up in a package inventory. This is what one of those copies looked like nine days later.

What Check Point disclosed about PHP2Shell

On 12 Nov 2026 Check Point published a security advisory confirming active exploitation of two issues used together.

The first is PHP2Shell, which Check Point describes as a vulnerability in PHP that also affects its Security Gateways. It covers CVE-2026-109618 and CVE-2026-109619, scored 9.2, and gives unauthenticated remote code execution through the firewall's web interfaces. The exposed component is the Mobile Access portal, the page remote users sign in to. Check Point lists every Security Gateway version from R80 to R82.20 with that portal, plus several Spark firewall releases, and tracks the fix in sk1000302. PHP itself patched the flaws on 21 Oct, and CISA added them to the Known Exploited Vulnerabilities catalog on 26 Oct, according to BleepingComputer.

The second is CVE-2026-123039, a local privilege escalation in the Gaia operating system, scored 7.8 and tracked in sk1000362. A root process called dynamic_server, part of the Dynamic CLI service, listens on a local socket that is created world writable, permissions 0777, and does not check who connects. Any process on the appliance that can reach the socket can have commands run as root. Check Point is clear that this needs local code execution first. PHP2Shell supplied it.

CVE-2026-123039 Critical
  • Product //

    Check Point Gaia: gateways, management and log servers, Multi-Domain, Maestro Orchestrator
  • Affected //

    R82.20, R82.10, R82, R81.20; R81.10 and R81 are end of support
  • Fixed in //

    Jumbo Hotfix R82.20 Take 14, R82.10 Take 52, R82 Take 132, R81.20 Take 174
  • Not affected //

    Spark firewalls, Smart-1 Cloud
  • CVSS //

    7.8, per Check Point
  • Exploited //

    Yes, chained with PHP2Shell; in CISA KEV since 13 Nov 2026

The timeline in Check Point's own words is the uncomfortable part. Exploitation "dates back to October 30 and was observed as active until November 4". Fixes for the gateway and management were available "one week before the attack, on October 23rd". The advisory says only a couple of dozen unpatched systems were compromised, that affected customers were notified in real time, and that in a limited number of cases the attack "also involved destructive activity and was ideologically motivated".

Check Point Research splits the activity into two waves. The first, on 30 Oct, targeted Israeli organizations and included destructive activity against some of them. The second, on 2 Nov, reached organizations across multiple countries and sectors, with no evidence of destructive activity. CISA added CVE-2026-123039 to the catalog on 13 Nov 2026.

The same Check Point advisory lists two more Gaia issues that are not known to be exploited: CVE-2026-107484, an authorization flaw in the Gaia Portal scored 7.3 that lets an authenticated user without admin rights perform admin operations, and CVE-2026-120490, scored 3.8. The November Jumbo Hotfix also carries the fix for CVE-2026-111497, the Brotli decompression flaw we covered in the copies your scanner misses. That is three upstream or platform bugs riding into one appliance release in a single month.

How PHP2Shell became root on a Check Point gateway

Check Point Research reconstructed the intrusion in detail. We describe it at the level of what each stage gave the attacker, not how to repeat it.

Figure 01

From a login page to a wiped hypervisor

  1. 01

    In //

    PHP2Shell on the Mobile Access portal, no login.

  2. 02

    Root //

    The open Dynamic CLI socket runs commands as root.

  3. 03

    Stay //

    Webshells, then the TSGA backdoor.

  4. 04

    Look //

    Config theft and an internal scan.

  5. 05

    Move //

    SSH to ESXi as its service account.

Result //

A wiper on the hypervisor, in some cases.

The firewall was the first host in the intrusion, not the last line of defense.

Stage one left traces in the portal itself. Check Point Research found PHP session files carrying the marker upload_progress_pwn under the Mobile Access temporary directory, and hundreds of segmentation faults in the portal's Apache and PHP workers. That crash pattern fits the class of bug: the PHP flaws corrupt memory that one request leaves behind for the next, and failed attempts kill workers.

Stage two turned the portal's service account into root through CVE-2026-123039. This is the step a well built appliance should have made hard. A web process that falls should land in a low privilege box. Here it landed next to a root daemon that accepted orders from anyone on the machine.

Stage three was about staying. The attackers wrote PHP webshells into the portal's template directory, then dropped a backdoor Check Point Research calls TSGA, running as /tmp/CPconfigdaemon so it looked like a Check Point component. Run as root, it relocated and relaunched under the name slot_CPconfigdaemon in a boot persistence directory, wrote a proxy component named TSProxy to /lib/libcpfwcommon.so, and listened on TCP port 11072. According to BleepingComputer, TSGA and TSProxy had previously turned up in investigations of attacks on Ivanti appliances, which fits what we said about edge VPNs after the patch: the implant outlives the fix.

Stage four was reconnaissance from a trusted position. TSGA can collect selected or full appliance configurations, and the attackers deployed the gogo scanner on the gateway. A firewall's configuration is a map of the network it protects: every subnet, every rule, every VPN peer.

Stage five is where it stopped being a firewall incident. The attackers moved over SSH to a VMware ESXi host, authenticating as vpxuser, the account vCenter uses to manage ESXi hosts. On the host they ran a wiper Check Point Research names destroy_v6 and left a Hebrew note named DREAM_ANNEXATION.txt with no payment demand.

Why the hotfix did not stop the wiper

There are three gaps in this story, and only one of them is about patching.

The first is the seven day window. The fix for the gateway shipped on 23 Oct and the first wave hit on 30 Oct. For an internet facing component with a public bug class and a CISA listing on 26 Oct, seven days was the whole budget. Teams that treat firewall Jumbo Hotfixes as a quarterly change window were outside it. We do not say that lightly: firewall changes break VPNs and the change board is right to be careful. But the decision to wait was a decision to run a known exploitable web application on the perimeter for weeks.

The second gap is inheritance. Most vulnerability programs track PHP on servers they know run PHP. Very few record that the firewall runs it too. When PHP2Shell was announced, the useful question was not which of our web servers run PHP, but which of our products embed it and when their vendor ships. Brotli raised the same question in October and the same Jumbo Hotfix answered both. A scanner that reads the gateway's banner sees a Check Point version, not a PHP version.

The third gap is the one the patch cannot close. A hotfix applied on 5 Nov to a gateway compromised on 30 Oct fixes the bug and leaves TSGA running. Check Point's indicators list says it plainly: absence of a match does not rule out compromise. Every gateway that had the Mobile Access portal exposed and sat unpatched at any point after 30 Oct needs a hunt, whatever its version is today.

Then the attribution. Check Point Research says TSGA and TSProxy are modified variants of PhiliKit, a toolkit ESET associated with the China nexus cluster UNC5221. It also says the destructive operation and the Hebrew note do not fully align with that actor, and that the evidence "does not allow us to assess with high confidence the true nature of this operation". It lays out three readings: the actor ran the whole campaign, acted only as an initial access broker, or took part in a false flag. We take that caution at face value. For a defender the answer does not change the work. The access was sold, shared or used for sabotage, and every one of those ends with someone else holding a root shell on your perimeter.

How to hunt for PHP2Shell compromise on Check Point

Start with the root socket. Check Point says the vulnerable daemon logs every command it receives to /var/log/dynamicCli.log. sk1000362 describes the checks: commands that did not arrive through the normal Clish wrapper may have been run as root through the socket, and the article lists search terms for exploitation such as reads of /etc/shadow, downloads and changes to authorized keys. On a patched system, lines reading "Rejected connection from unauthorized uid" mean someone tried after the fix. After the hotfix, the socket should show permissions 0660, owned by root and the config group, not 0777.

Then the portal. Look for PHP session files containing upload_progress_pwn, unexplained template files in the portal's template directory, and a burst of segmentation faults in the portal's web workers from 30 Oct onward.

Then the host. Check Point's indicators include a process or binary named CPconfigdaemon in /tmp, the slot_CPconfigdaemon copy, a library at /lib/libcpfwcommon.so, a listener on TCP port 11072, and scanner or wiper tooling under /tmp. If the appliance ships host events anywhere, this rule is a starting point.

Figure 02

TSGA and TSProxy artifacts on a Check Point gateway

Sigma

title: Check Point gateway TSGA or TSProxy artifacts, PHP2Shell campaign
status: experimental
description: File paths Check Point Research lists for the TSGA backdoor and TSProxy on gateways compromised from 30 Oct 2026
references:
  - https://research.checkpoint.com/2026/php2shell-exploitation-on-check-point-gateway/
logsource:
  product: linux
  category: file_event
detection:
  selection_tsga:
    TargetFilename|endswith:
      - '/tmp/CPconfigdaemon'
      - '/slot_CPconfigdaemon'
  selection_proxy:
    TargetFilename: '/lib/libcpfwcommon.so'
  selection_note:
    TargetFilename|endswith: '/DREAM_ANNEXATION.txt'
  condition: 1 of selection_*
falsepositives:
  - Unlikely; none of these names is a shipped Check Point file per the research
level: critical
A starting point to tune against your own telemetry, not a production rule.

Then the network. Block and search for the two scanning sources Check Point published, 130.94.19[.]204 and 130.94.19[.]168, in the portal's access logs. Look for SSH sessions that start on the gateway and end anywhere inside, above all on ESXi hosts. A firewall has no business logging in to a hypervisor, and a vpxuser login from anything other than vCenter is a finding on its own.

Then the gaps. Check Point lists unexplained holes in firewall traffic logs, especially after the logging daemon crashed, as an indicator. The device that writes the logs is the device that was owned, so compare against what the next hop saw.

A hit is not a cleanup job. The attacker held root on a device that stores VPN secrets, certificates, administrator accounts and the configuration of the whole network. Rebuild the gateway from known good media, rotate everything it held, and rotate the vCenter and ESXi credentials it could reach. Do this on the management server too, since CVE-2026-123039 affects management servers as well.

How we test a firewall's own attack surface

On an external test we treat the firewall as a set of services, not a perimeter. We inventory every interface it answers on: the VPN, the user portal, the administration portal and anything a feature switch left on. We read what each one discloses without logging in, from version hints to error pages. Our exposed admin and debug endpoint checks show the method, and atarus-recon on our open source page covers the wider external inventory. A user portal that most staff never use is a common finding. So is an administration page that should only answer on a management network.

On an internal or assumed breach engagement we ask the question this campaign answered. If the gateway falls, what does it reach? We map the routes and credentials from the firewall to management networks, hypervisors, directory servers and backup systems. Can it open SSH to the ESXi management interface? Does vCenter's service account work from anywhere but vCenter? Are the gateway's administrator credentials reused on other infrastructure? Our remote admin service checks cover how we read SSH exposure on the inside without trying a login.

We also review patch practice for appliances as a finding in its own right. We ask how long the last three firewall Jumbo Hotfixes took from release to install, and who decided. We ask whether anyone tracks the third party components inside the appliances, PHP among them. We ask whether the gateway sends host telemetry anywhere, or only traffic logs. A device that only reports on traffic will not report on itself.

What we do not do is run the exploit against a production firewall. A memory corruption bug that crashes web workers on failure is not something we fire at a client's only VPN gateway to prove a point. We identify the exposure from the outside, confirm the version and hotfix state with the client, and test the blast radius from a foothold we are given.

What is still coming

Expect more vendors to name PHP2Shell. Check Point is the first large appliance maker to confirm exploitation through it, but PHP sits inside many products with web administration and user portals. Each one ships its own fix on its own schedule. Ask your vendors directly whether their products embed an affected PHP build and when the fix lands.

Expect the end of support versions to be targeted. R81.10 and R81 are affected by CVE-2026-123039 and are marked end of support in sk1000362 with no fixed take listed. Any gateway still on those releases has no fix for this chain except an upgrade.

Expect the implants to outlast the headlines. TSGA has now been seen on Ivanti and Check Point gear. A backdoor that survives on two vendors' appliances will be looked for on others. Hunts run this month should be repeated.

Expect the question of who did it to stay open. Check Point Research was careful, and we will not get ahead of it. The work for defenders is the same either way.

What to do today

  1. Install the November Jumbo Hotfix on every gateway and management server. Start with any gateway running the Mobile Access portal. The fixed takes are R82.20 Take 14, R82.10 Take 52, R82 Take 132 and R81.20 Take 174; R81.10 and R81 need an upgrade.

  2. Where you cannot patch yet, apply the sk1000362 workaround and close the portal. Stopping the Dynamic CLI process breaks some commands; turning off or restricting a Mobile Access portal you do not need removes the way in.

  3. Hunt on every gateway that was exposed after 30 Oct, patched or not. Read /var/log/dynamicCli.log, check the socket permissions, look for the TSGA and TSProxy files and the upload_progress_pwn session files, and use the Sigma rule above as a starting point.

  4. Cut the firewall's path to your hypervisors. Block SSH from gateways to ESXi and vCenter, and alert on any vpxuser login that does not come from vCenter.

  5. On a hit, rebuild and rotate. Rebuild the gateway, then rotate its VPN secrets, certificates and admin accounts, and the vCenter and ESXi credentials it could reach.

  6. Add embedded runtimes to your inventory. Record which appliances carry PHP, Node.js and similar components, and track their vendor fixes with the same deadline as the upstream fix.