# The Cisco Expressway attack was rewritten with the help of AI

> F6 investigates the first real abuse of CVE-2026-76466, a 9.8 severity flaw in Cisco Expressway, with Python scripts that the attacker tweaked during the assault.
> By José Maldonado · Cybersecurity · 2026-11-20

A remote access gateway that Cisco classified as critical went a month without anyone touching it. The vulnerability CVE-2026-76466, 9.8 out of 10 on the CVSS scale, lives in Cisco Expressway, the equipment through which a company gives its employees access to their voice and video meetings. [Cisco's advisory](https://sec.cloudapps.cisco.com/security/center/content/CiscoSecurityAdvisory/cisco-sa-vcs-static-cred-hr7EWZ8t) was published on October 21 with an uncomfortable warning for those who administer these systems: there is no way to mitigate it other than updating.

The Russian cybersecurity firm F6 reported on November 18 what came next. Its team investigated the first known case of real exploitation, at an organization in Russia with two Expressway X15.5.3 appliances exposed to the internet and without the patch applied. [F6's note](https://www.f6.ru/media-center/press-releases/cisco-cve-2026-76466/) reconstructs the full path, from those two perimeter machines to two cluster nodes that lived on the internal network, out of reach of any scan from outside.

What is new is not the flaw, but the attacker's way of working. The Python scripts used to move through the network changed while the assault was underway, and the code carried comments and descriptions of fixes that, in F6's words, indicate the use of an AI assistant. [That detail turns one more incident into a sample of something larger](https://www.securityweek.com/hackers-exploit-cisco-expressway-vulnerability-soon-after-poc-release/): writing intrusion tools has become cheap, and that shortcut leaves traces in the code's own text.

## A credentials flaw that only reaches clusters

Cisco's advisory describes the error bluntly. The Expressway database service uses persistent cookies and does not properly validate client certificates when TLS verification mode is enabled for the cluster.

Anyone with those cookies reaches the service, launches RPC commands (remote procedure calls, the mechanism by which one process entrusts work to another) and ends up with root privileges, the superuser of the operating system.

The cluster condition limits the scope, but does not make it small. A standalone Expressway is not vulnerable; a pair of nodes that talk to each other to share state is, and that is precisely the configuration of anyone seeking high availability. The flaw falls into the CWE-798 family, use of credentials embedded in the product, and that explains why the vendor offers no workarounds: no configuration setting can cover a trust that the software itself installs and the customer cannot change.

The advisory also says whom to thank for the finding. Stefan Aleksic, of the financial firm Jane Street, reported the problem to Cisco, and the fix arrived in Expressway X15.5.20. Equipment on X14 or earlier branches must migrate to a supported version, because the advisory does not publish a patch for those trains.

Data point

Value

Identifier

CVE-2026-76466, CWE-798

Severity

CVSS 9.8 out of 10

Published

October 21, 2026

Affects

Expressway-C and Expressway-E, virtual and appliance, cluster only

Fixed in

X15.5.20; X14 and earlier, migrate

Service ports

TCP 4371 and 4372

The calendar is the data point that weighs most in practice. Between the October 21 advisory and the November intrusion there is less than a month, and in that gap there were only two ways out: update to X15.5.20 or hope no one looked. Exploitation arrived before the second.

## Two machines on the internet and two nodes inside the network

F6's reconstruction begins at the network edge. Two Expressway appliances on version X15.5.3, accessible from the internet, were compromised, and from one of them the attacker jumped to two more nodes of the same cluster that had no direct route to the public network. In a cluster the nodes trust each other by design, and that trust is what turns one compromised machine into a key that opens the others.

With root in hand, the next step was to prepare tools. F6 describes custom Python scripts to scan the network, locate SMB resources (the protocol with which Windows shares files and folders), talk to SQL Server and WinRM (Windows' remote administration channel) and inspect GitLab repositories. The objective recorded in the note was to extract credentials and other sensitive data from the internal infrastructure.

The trigger was the proof of concept. Since someone published [an exploit on GitHub](https://github.com/dinosn/cve-2026-76466), anyone with a clustered Expressway can check whether theirs responds, and the repository itself includes a check mode that only authenticates and does not execute commands, designed to inventory without attacking.

F6 places the assault shortly after that publication, while Cisco's advisory, unreviewed since October, still states that the vendor is unaware of malicious use of the flaw. The distance between the two documents sums up the week.

## An assistant's comments in the attacker's code

The detail that sets this case apart from other similar ones is in the text of the scripts. F6 found comments and correction notes that, in its judgment, can only be explained by the use of an AI assistant during the assault. The reading fits with what is already seen in ordinary development: whoever asks a model to fix a function gets back a comment justifying the new line, and that trace stays attached to the file if no one cleans it before using it.

For the attacker, the gain is time. Modifying a tool in the middle of an intrusion forces you to stop, test and go back in; an assistant shortens that cycle to minutes and allows the script to be adapted to the environment that appears, something that previously distinguished the most skilled teams. The same capability has its defensive side, and [OpenAI's work on 948 privilege escalation flaws in the Linux kernel](https://observatorioblockchain.com/en/ai/openai-detecta-948-fallos-linux-tomar-control/) shows that models also serve to review third-party code at scale.

It is worth separating two things that the headline conflates. That an attacker uses AI does not change the severity of the flaw; it changes the pace at which it is exploited and the number of people capable of trying. And detection based on code style works as an indication in a forensic report, not as a control you can deploy in production, because the next person who carefully rewrites the script erases the clue.

## The trace in Erlang and the ports worth closing

The vulnerable service is part of the Erlang cluster, the environment with which Cisco built Expressway's distributed database. F6 points to a specific line in the system log, erlang.log, as the cleanest sign that someone used the flaw: an entry \`nodeup,'wombat@127.0.0.1'\`.

That node is not a legitimate cluster peer, so its presence in the log makes it necessary to review everything else, from local accounts to the processes that started afterwards.

The note also publishes two IP addresses associated with the attack, 66.179.31.32 and 18.130.47.116, and the pattern of paths where the scripts appeared, under /mnt/harddisk/ with a directory beginning with the letter i and four digits. These are indicators of this incident and not a universal blocklist. F6 does not identify the victim or attribute the attack to a known group, and this article does not either.

The interim mitigation also comes from F6, because Cisco offers none. Blocking connections to TCP ports 4371 and 4372 from any address other than a trusted node of the cluster keeps out the intruder arriving from the internet or from a neighboring segment.

It is a network filter and not a fix, so it only buys the weeks it takes to update, the same kind of emergency measure that has been repeated this week with [the MPTCP flaw that remains unpatched in RHEL](https://observatorioblockchain.com/en/cybersecurity/cve-2026-98432-mptcp-linux-sin-parche/).

## A collaboration gateway that almost no one watches

Expressway is that kind of infrastructure that is installed once and forgotten: it provides remote access to corporate voice and video, generates no revenue and rarely makes it into the quarterly security review. The result is a fleet of machines on old versions, grouped in clusters, with ports open to the internet and an internal trust that an attacker turns into lateral movement. In a week with [1,094 vulnerabilities fixed by Microsoft](https://observatorioblockchain.com/en/cybersecurity/patch-tuesday-noviembre-2026-outlook-cero-dia/), the voice gateway is on no one's list.

It remains to be seen whether Cisco revises the October advisory with the exploitation notice and whether more organizations appear with the wombat entry in their logs. Meanwhile, the useful question for any team is one of inventory: how many clustered Expressways there are, on what version and from where ports 4371 and 4372 are reachable. That list decides whether this week's incident is someone else's anecdote or a pending warning.
