The switch in a community network closet is the least examined device in the stack. This story is about what walks in through its least examined interface: the management web UI, reachable before login, running a CGI handler on a MIPS core that trusts a counter.
The August event¶
GreyNoise reconstructed a single malicious actor — one infrastructure footprint tracked back to early June, the same or related to the "Red Heron" activity Acronis published in September — and put a timestamp on the switch campaign: on or about August 17, the actor exploited and exfiltrated data from 996 ZyXEL GS1900 Smart Managed Switches in 48 countries.
What went out, per GreyNoise: device configurations, root-level credentials (hashed), and networking information. The stolen credential hashes still matter: 564 of the 996 victims had factory default credentials, which means the exfil handed the actor valid root logins on more than half the fleet and, on the rest, cracked-or-reused password material tied to a real network. Where they were: Italy 133, United States 129, Taiwan 123, France 90, South Korea 69, Netherlands 66, Czech Republic 49, United Kingdom 33 — and then smaller counts through 48 countries. Small deployments everywhere. The distribution is the thesis: this was not a campaign against data centers, it was harvesting the long tail — exactly the boxes this site's readers run.
The tooling was a Python script obfuscated with the commercial PyArmor tool — deobfuscated in part because the actor left a runtime pinning it to the 2021-era version 6.7.5. The script's one purpose was exploiting the GS1900 pre-auth handler. It targeted /cgi-bin/dispatcher.cgi — the entry point on the Boa web server embedded in the switch firmware's MIPS core.
The mechanics¶
From Binarly's analysis (researchers Yegor Vasilenko and Fernando Mercês, who built on GreyNoise and Acronis rather than duplicating them): the exploit needs one crafted HTTP POST in GOT mode — deterministic, one request, done — or a stack-mode average of about 2,048 attempts against ASLR. The command payload is arbitrary; GreyNoise captured the actor's actual choice: a busybox telnetd one-liner to hold the box, then a TFTP pull of a collector (tftp -gr c ... ; /bin/sh /1) whose final act stages /tmp/info into the web server's own served directory, /home/web/tmp/info.txt, where it could be retrieved like any web asset.
The bug itself is an out-of-bounds write — a crafted request reaches an offset controlled by a parameter counter the loop never validates, so the copy lands wherever 1024 × counter points. Binarly believes the exploit is LLM-generated; their counter-move was the same tooling, standing up emulated firmware environments to test the patches. That symmetry is now a fact of the field and it is worth saying plainly: the tool that wrote the exploit is the tool that can verify the patch.
The patch that did not work¶
Here is the part that turns a bug story into a ledger story.
- June 16, 2026 — Zyxel ships firmware 2.90(Axxx.2)C0 across the line, closing CVE-2026-7273, which NVD describes as a stack-based buffer overflow in the CGI program.
- September 21, 2026 — GreyNoise publishes the August campaign and ties the exploit to CVE-2026-7273. CISA adds CVE-2026-7273 to the KEV catalog the same day, with the standard recommendation: upgrade your firmware.
- September 28, 2026 — Binarly, having decompiled the actor's exploit, confirms that all ten fixed releases addressing CVE-2026-7273 remain vulnerable to the very same exploit, working without changes. The patched code had put two checks in: a
sprintf-to-snprintfswap inmain(protects a different, stack buffer the exploit never touches) and an equality test inside the parameter loop — but the test runs after the counter is already corrupted, so the comparison fails, execution falls through to the branch that processes the last parameter, and the 141st parameter of the payload gets copied intoreqParameters[1024 × counter]anyway. The fix shipped in June fixed a buffer overflow in the same file. Not the path the exploit walks. - September 29 — Binarly reports to Zyxel. The vendor's response is honest but reads like taxonomy defending itself: CVE-2026-7273 was, Zyxel says, about the
reqParameterscounter overflow all along — despite their own advisory text calling it a stack overflow — and the missing parameter-size check Binarly found is a separate vulnerability because it is a separate missing check. That new check-less write gets the identifier CVE-2026-107441. - October 11 — Zyxel confirms and reserves the CVE.
- October 22 — Real fix ships, 2.90(AAxx.3)C0: copies capped at 511 bytes per field, and the parameter count checked before the copy into the global buffer.
- November 3 — The advisory goes public, CVSS 8.8, "a LAN-based, unauthenticated attacker" and "potentially execute OS commands," acknowledgment to Binarly. Canada's Cyber Center posted its own advisory within days, and heise reported attacks still observed against the unpatched base.
Count the distance: the exploited bug existed when the June patch shipped, the KEV deadline of September 21 pointed thousands of admins at firmware that did not stop the actor, and the machine-readable truth — patch-3 firmware — landed October 22, eighteen days before anyone outside the PSIRT could read about it.
What this site has seen already, now in one place¶
The same week Keycloak shipped three fixes that did not hold inside one tag, this site wrote: assume the previous tag's patch list is provisional, not settled. GS1900 is that sentence with a firmware table and a victim count. The class repeats: a vendor ships a fix, a deadline mechanism (KEV) certifies urgency against the fix, and the exploit's actual path is elsewhere in the file. Version numbers are a filing system — CVSS is a filing system too, and NVD's description of CVE-2026-7273 named the wrong memory region for four months while the real thing sat in .bss.
The difference worth marking: Keycloak's failed fixes were caught by the vendor's own process between releases. Zyxel's failed fix shipped to every customer and stayed the official answer for two additional months of active exploitation, and this was discoverable only because one lab read the attacker's bytecode against the patched binary. For appliances you cannot test, the honest rule is: a fix landed when the release notes describe the mechanism patched — not when a CVE row on a catalog says closed. On this incident, CISA's own catalog entry was the deadline on the wrong door.
The firmware table¶
Every GS1900 model variant carries its own build tag. Check your version string against this from Zyxel's advisory:
| Model | You are patched at or above | | --- | --- | | GS1900-8 | 2.90(AAHH.3)C0 | | GS1900-8HP | 2.90(AAHI.3)C0 | | GS1900-10HP | 2.90(AAZI.3)C0 | | GS1900-16 | 2.90(AAHJ.3)C0 | | GS1900-24 | 2.90(AAHL.3)C0 | | GS1900-24E | 2.90(AAHK.3)C0 | | GS1900-24EP | 2.90(ABTO.3)C0 | | GS1900-24HPv2 | 2.90(ABTP.3)C0 | | GS1900-48 | 2.90(AAHN.3)C0 | | GS1900-48HPv2 | 2.90(ABTQ.3)C0 |
Patches exist only for models still inside Zyxel's support window — the table above. Anything else in your rack with GS1900 printed on it is, by the vendor's own partition of the world, out of support and unpatchable. Retire it or wall it off; it is running the June-and-before firmware that this actor demonstrably empties.
The operator checklist¶
For a community network, fire hall, or small office running GS1900-class hardware:
- Inventory. Every managed switch, its firmware string, its model. The version is on the management UI's system page and in its config dump.
- Patch in-support boxes to the .3C0 builds above — this is the CVE-2026-107441 fix, the one that actually matters.
- Treat August 17 through October 22 as burned ground for anything left on older firmware: if a GS1900 served a user-facing VLAN with its management UI enabled during that window, assume its configuration and root credential hashes are known. Rotate every credential that machine knew — the switch admin password, yes, but also anything it saw: management VLAN addresses, upstream router credentials, SNMP strings, and every other device whose password reuse touched it.
- Hunt. From GreyNoise's capture: a file staged at
/home/web/tmp/info.txt, TFTP requests leaving the switch to port 6969,telnetdlisteners you did not start, configuration diffs you did not make. The web server's own directory serving the collection is the tell to grep for in any mirror you keep. - Pull management off the data plane. The bug is pre-auth, but it needs the management UI reachable. A dedicated management VLAN, an ACL that names one admin jump host, and management interfaces disabled on access ports convert this class of bug from "empty the fleet" to "log noise from a jump box." Whatever your switch vendor calls that feature, use it.
- Default credentials, one more time. Over half the harvested devices held factory passwords. The pre-auth bug did not need them — the next thing on the network will anyway.
The closing honest note¶
Zyxel's handling improved at the end: fast reservation-to-patch window (Sept 29 → Oct 22 is twenty-three days), and an advisory that names the reporter and the class precisely. The failure was upstream of the vendor: a June "fix" that did not close the exploited path, and a catalog deadline that certified it. That chain — vendor ships, catalog stamps, exploit ignores — is the reason this site keeps writing operator-facing timelines instead of CVE counts. The community networks this field builds run on hardware like this, funded by bake sales and grant cycles, maintained by one person with a laptop on an evenings-and-weekends SLA. Those operators deserve timelines, firmware tables, and hunt patterns — and the standing permission to trust a release note that describes a mechanism over a CVE row that describes a status.
And tonight's companion note: the other thing every such network runs is sshd, and it has a bug of its own this week.