996

GS1900 switches GreyNoise says were exploited and had data taken

564

of those victims still on factory default credentials

10

patched firmware releases the exploit still worked against, per Binarly

31

days from the KEV listing to the firmware that stops the exploit

A patched CVE is not a patched device

For five weeks this autumn, a Zyxel GS1900 switch could be fully patched against the vulnerability on CISA's exploited list and still fall to the exploit that put it there. The June firmware fixed a bug. It did not fix the one the attack tool was using. Owners who met the federal deadline had done everything the paperwork asked, and their switches stayed open.

That is the point we want buyers and defenders to take from CVE-2026-107441. A CVE ID on an exploited list is a claim about a bug. It is not proof that your device resists the attack that is actually happening. Those two things usually line up, which is why vulnerability programs are built on CVE IDs and version numbers. When they do not line up, nothing in a normal patch report will tell you.

The second point is older and plainer. Every one of these attacks needed someone to reach the switch's web management page. A GS1900 is a small office switch. Its management interface belongs on a network nobody else can touch, and on 996 switches in 48 countries it was not. More than half of those victims were still on factory default credentials, which tells you how much attention those management planes were getting.

We wrote in June that attackers spend whole weeks walking in through everything except the core applications. Switches, gateways and appliances are where that happens, and the GS1900 story adds a sharper lesson: on this class of device, the patch you can see on a version string deserves less trust than the exposure you can see on a network map.

What is Zyxel CVE-2026-107441?

CVE-2026-107441 is an out-of-bounds write in the CGI program behind the web management interface of Zyxel GS1900 series switches. Zyxel's advisory, published on 3 Nov 2026, says it could allow a LAN-based, unauthenticated attacker to potentially execute OS commands via a crafted HTTP request. It lists ten affected models, from the GS1900-8 to the GS1900-48HPv2, and gives a fixed firmware build for each: the release ending in .3, such as 2.90(AAHL.3)C0 for the GS1900-24.

The research firm Binarly reported the bug. Its own advisory rates it 9.8 under CVSS 3.1 with a network attack vector, and describes an unauthenticated remote attacker running arbitrary commands as root. Zyxel's wording is narrower: LAN-based. The difference matters less than it looks. A web interface reachable from the internet is, for practical purposes, on the attacker's LAN.

CVE-2026-107441 Critical
  • Product //

    Zyxel GS1900 series switches, web management CGI
  • Affected //

    ten models, firmware 2.90(xxxx.2)C0 and earlier
  • Fixed in //

    2.90(xxxx.3)C0 for each model, released 22 Oct 2026
  • Authentication //

    none
  • CVSS //

    9.8 (3.1, Binarly); Zyxel describes a LAN-based attacker
  • Exploit in use //

    works on the June fix, per Binarly

The history is what makes this one worth a long read.

On 16 Jun 2026 Zyxel published an advisory for CVE-2026-7273, a pre-authentication buffer overflow in the same switches, and shipped firmware ending in .2 to fix it. On or about 17 Aug, according to GreyNoise, an attacker exploited GS1900 switches and took their configuration, network information and hashed root credentials from 996 devices across 48 countries. GreyNoise says 564 of those victims had factory default credentials. The largest numbers were in Italy, the United States and Taiwan.

GreyNoise published on 21 Sep and tied the activity to CVE-2026-7273. CISA added CVE-2026-7273 to its Known Exploited Vulnerabilities catalog the same day, based on evidence of active exploitation, and gave federal civilian agencies until 24 Sep to act. Everyone did the sensible thing: find the GS1900s, install the June firmware, close the ticket.

Binarly was looking at the same attacker from another direction. While studying malware infrastructure linked to a campaign that also hit Gitea servers, it found the attacker's Zyxel exploit sitting in an open directory. On 28 Sep it confirmed that the exploit worked against all ten firmware releases that fixed CVE-2026-7273. It told Zyxel on 29 Sep. Zyxel confirmed a new vulnerability on 11 Oct, released the .3 firmware on 22 Oct and published the advisory on 3 Nov.

From the KEV listing on 21 Sep to the firmware that stops the exploit on 22 Oct is 31 days. For that whole month, a GS1900 that had been patched exactly as instructed was still open to the tool in use.

Two careful points. First, Zyxel's advisory for CVE-2026-107441 does not say the new bug was exploited, and BleepingComputer notes it is unclear whether the attacker built the exploit to beat the June fix or simply hit a related bug the fix missed. Second, nobody has said that patched switches were among the 996 victims. What Binarly showed is that the exploit worked against them. When we checked the KEV catalog release of 4 Nov, CVE-2026-107441 was not listed; CVE-2026-7273 was.

How a June patch missed an August exploit

The mechanism explains why the June patch looked right and was not, so we describe it at the level of what each step achieves, not how to do it.

The switch's web interface parses every request into a fixed table of parameters held in memory. The table has room for a set number of entries of a set size. Right after the table in memory sits the counter that records how many parameters have been read. The parser did not check either the size of each parameter or the count.

Binarly's analysis shows the June firmware added a check on the count. It did not add a check on the size. An oversized parameter at the right position spills past the end of the table and overwrites the counter itself. Once the counter holds an attacker's value, the parser's next write lands wherever that value points, before the new count check gets a chance to object. From there the exploit redirects a routine function call to run a command of the attacker's choosing, as root, with no login.

Zyxel told Binarly that CVE-2026-7273 was the missing count check, even though its June advisory described a stack overflow, and treated the missing size check as a separate flaw. That is how one exploit ended up spread across two CVE IDs, with the public record pointing at the first.

Figure 01

One exploit, two CVE IDs

  1. 01

    Reach //

    The attacker reaches the switch's web management page.

  2. 02

    Fingerprint //

    The exploit reads the firmware build, June fix included.

  3. 03

    Overflow //

    An oversized parameter overwrites the parameter counter.

  4. 04

    Execute //

    The corrupted counter becomes a command run as root.

  5. 05

    Collect //

    Configuration, network details and hashed root credentials leave.

Result //

The June patch was installed, the check it added was skipped.

The fix closed the bug the advisory named. The exploit used the bug beside it.

The detail that should worry anyone running a patch program is in step two. Binarly found that the exploit's built-in version table included the June build for the GS1900-24, the firmware released to fix CVE-2026-7273. Somebody had tested the attack against the patch. The attacker knew the fix did not work before the defenders did.

Binarly says it believes the exploit was generated with a language model, and GreyNoise made a similar observation about other code from the same attacker. Neither presents that as certain, and neither do we. It does not change the defence. A tool that fingerprints the firmware and carries a table of targets works the same however it was written.

Why a KEV deadline can be met and still miss

We have no quarrel with CISA here. The KEV listing was right on the evidence available on 21 Sep: GreyNoise had tied the exploitation to CVE-2026-7273, and a fix existed. The process worked as designed. The design has a gap, and this case shows it more cleanly than most.

Vulnerability management runs on identifiers. A scanner reports a CVE, a feed says it is exploited, a patch is mapped to it, a version check confirms the patch, the ticket closes. Every link in that chain is a claim made by somebody: the researcher who tied the exploit to the CVE, the vendor who wrote the fix, the scanner that reads the version string. When the first claim is slightly wrong, every later step faithfully carries the error forward, and the dashboard shows green.

That is why we keep saying that a version string is evidence, not proof. We made the same argument about PHP builds that report old version numbers after vendor backports, and about Ivanti appliances that were patched but not clean. The GS1900 is the reverse of the PHP case. There, a fixed build could look vulnerable. Here, a vulnerable build looked fixed.

What breaks the chain is a question the identifier cannot answer: does this device resist the attack that is happening? For most teams the honest answer is that they cannot test that themselves, and should not try on production network gear. What they can do is reduce how much rides on the answer. A switch whose management page cannot be reached from a user network, a guest network or the internet was safe from this exploit in June, in September and in October, whatever its firmware said.

There is a second gap. Zyxel's advisory says it released patches for models still within their vulnerability support period, and that products not listed are unaffected. Older GS1900 hardware that has left support appears in no table and gets no fix. If you own one, the exposure decision is the only control you have left.

How to detect GS1900 exploitation

Small switches rarely send useful logs, so detection here works best from the network around them.

Start with who talks to the management interface. List the addresses your GS1900 switches answer on and check firewall, NetFlow or switch port logs for HTTP or HTTPS sessions to them from anywhere other than your management hosts. Any session from a user subnet, a guest network or the internet is either a misconfiguration or an attack, and both need fixing.

Then watch what the switches start themselves. A switch has little reason to open outbound connections beyond NTP, syslog and perhaps a firmware server you named. GreyNoise reports that after exploitation the attacker pulled a collection script onto the device over TFTP from its own infrastructure. A switch that initiates TFTP, or any connection, to an address outside your network is a strong signal. Binarly also lists command and control hostnames under the domain 981666.xyz; a DNS query for that domain from anything on your network deserves an incident ticket.

Figure 02

Switch management address starts an outbound connection to the internet

Sigma

title: Network switch initiates outbound connection to an external address
status: experimental
description: A GS1900 or similar managed switch opening a connection to a non internal address, such as the TFTP pull GreyNoise observed after GS1900 exploitation. Replace the address list with your switch management range.
logsource:
  category: firewall
detection:
  switches:
    src_ip|cidr:
      - '10.0.250.0/24'
  internal:
    dst_ip|cidr:
      - '10.0.0.0/8'
      - '172.16.0.0/12'
      - '192.168.0.0/16'
  condition: switches and not internal
fields:
  - src_ip
  - dst_ip
  - dst_port
falsepositives:
  - Firmware or time servers you configured on purpose; list them in the internal selection
level: high
A starting point to tune against your own telemetry, not a production rule. Pair it with a DNS rule for the 981666.xyz domain from Binarly's indicators.

If you find a hit, assume the switch configuration and its root password hash are in the attacker's hands. Replace the firmware with the .3 build, reset every credential the switch holds, and check whether the same password is used anywhere else, because GreyNoise says the collected data included hashed root credentials and a hash of a reused password is a key to more than one door. Review the configuration for changes you did not make, including new accounts, changed SNMP communities and mirrored ports. Binarly publishes file hashes and YARA rules for the JITTERLY backdoor and SIXZUT rootkit seen with this attacker on Linux servers; those apply to the servers the same campaign touched, not to the switch.

How we test network gear on an engagement

On internal tests, the network gear is usually the least loved part of the estate. Core switches get attention. The eight port switch in a branch office, the PoE switch behind the reception desk and the unit a contractor installed for the cameras often do not. Our internal network first hour starts from a foothold and an ordinary account, and one of the first things we map from there is which management interfaces answer us.

What we look for is reachability before version. From each network position in scope, we record which switches, access points and appliances expose a web interface, SSH, Telnet or SNMP to us. A management page that answers from a user VLAN is a finding on its own, whatever firmware is behind it, because it puts the device's security entirely in the hands of its own code. This bug is a good example of why that code does not deserve the trust.

Then we check the cheap things an attacker checks: factory default credentials, default SNMP communities, and configuration backups left on shares. GreyNoise's figure of 564 victims on default credentials is in line with what we see. We confirm a default credential by logging in and recording that we could, and we stop there.

We do not run memory corruption exploits against a client's production switches to prove a point. A crash takes down a floor of users, and the finding stands without it: a vulnerable build plus a reachable management page is the exposure. Our guide to proving network impact without causing damage covers how we show what a position could reach while leaving it as we found it. If a client wants a bypass like this one demonstrated, we agree a lab unit with them first.

On external tests, we look for the same interfaces from the internet. Our external recon method and the atarus-recon tool on our open source page find what an organisation exposes from its own domains and ranges. A switch login page on a public address is one of the findings we would rather report on day one than read about in a GreyNoise post.

What is still coming

Expect CVE-2026-107441 on the KEV list. The exploit is public knowledge, Binarly says it plans to publish a minimised proof of concept, and the attacker already has a working tool. Whether or not CISA lists it, treat it as exploited.

Expect more scanning of GS1900 management pages. The attacker's tool carries a table of firmware builds going back to 2015. Every switch that installed the June fix and stopped there is in that table. So is every switch that never installed anything.

Expect more patch bypass stories on network gear. Vendors of small appliances often fix exactly the input in the report, and attackers have every reason to test the fix. This case is unusual only because a researcher found the attacker's own exploit and could compare it with the patch. Most of the time nobody looks.

And expect the old hardware to stay exposed. Zyxel's table covers models in their support period. Anything outside it stays vulnerable for as long as it stays plugged in, and the only fix there is to take its management page off the network or replace it.

What to do today

  1. Take switch management off every network an attacker can reach. Restrict the GS1900 web interface to a management VLAN or named admin hosts. This stops the exploit whatever the firmware says, and it is the only control for models out of support.

  2. Install the .3 firmware for each GS1900 model. The June .2 build is not enough. Check the build on the device itself, not in an asset record.

  3. Change every default credential on network gear. More than half of the 996 victims were on factory defaults. Rotate the root password and SNMP communities on any switch that may have been reached.

  4. Hunt for switches talking to the internet. Outbound TFTP or any external connection from a switch address is a strong signal. The Sigma rule above is a starting point, and the 981666.xyz domain is worth a DNS search.

  5. Close KEV tickets on behaviour, not only on version. For exploited bugs on edge and network devices, record whether the device is still reachable by the attack, not just whether the patch is installed.