A Chat Message Could Swap a Mattermost Plugin
An ordinary user could make an administrator's browser replace a Mattermost plugin, and plugins run code on the server. Here is the chain and the fix.

The interesting part of CVE-2026-93851 is not the severity score. It is who sends the malicious request.
Mattermost published MMSA-2026-00742 on 28 October 2026 describing a path traversal in the Agents plugin's mention reminder. The advisory's own wording is precise: a crafted post could cause the mention reminder to send "an unintended request from an administrator's session" because the component did not validate the target post ID. The consequence, per the CVE record, is replacement of an installed plugin and code execution on the server.
Read that again as a sequence. An attacker who can post in a channel an administrator can see writes a message. The administrator clicks a mention reminder. The administrator's own browser issues a request the administrator never intended. A plugin gets replaced. Code runs on the server.
The chain, step by step
Nothing in that chain is individually unusual, which is what makes it worth understanding.
- The attacker needs post permission, not administrator rights. The advisory's precondition is an authenticated user who can post in a channel visible to a system administrator. On a busy self-hosted instance that is most of the company.
- The payload is a post containing a forged mention reminder. The reminder is a UI affordance people trust — it is how Mattermost tells you someone mentioned you.
- The mention reminder component trusted the post ID it was given. It did not check that the ID actually referred to a post the viewer was entitled to act on, so the value could be shaped to point somewhere else. That is the CWE-22 path traversal: a trusted-looking identifier used without a directory-equivalent boundary check.
- The administrator's session performed the action. The request arrived with a valid administrator session cookie, from the administrator's own browser, over the administrator's own connection. From the server's point of view it looks exactly like an administrator doing their job.
- The target of that request was plugin replacement. In Mattermost, plugins are not themes. They extend the server.
Why the administrator's browser is the interesting part
Most traversal bugs are server-side: an attacker sends ../ to your web server and the server does the wrong thing. Here the server is not the confused deputy. The browser is.
That distinction matters for defenders because the usual controls are pointed at the wrong place. A web application firewall inspecting inbound requests sees nothing anomalous — the malicious request is outbound from a trusted browser to the application, carrying a valid session. Server-side request validation never runs on a path the attacker supplied, because the attacker never supplied one to the server directly. What the attacker supplied was a chat message.
This is the forced-browse family, and it has an uncomfortable property: the more privileged your administrators are, the more damage a single click delivers. An administrator's browser is the highest-value session on a self-hosted collaboration server, and it is also the one that renders untrusted content by design. Chat is the product.
Plugin replacement is code execution by design
Mattermost's own developer documentation describes plugins as a way to extend "both the Server and Web/Desktop Apps". That is the feature, not a bug. A server plugin is a deployed binary with the server's privileges; installing or replacing one is a code-deployment operation, which is exactly why it normally requires administrator rights and an explicit upload.
So the severity is not inflated by the CVSS calculator. The chain ends where it should: plugin replacement is arbitrary code execution, and the vulnerability provides a way to reach that operation through a link a privileged user clicks.
The score recorded for CVE-2026-93851 is 8.0 HIGH on CVSS 3.1, with an exploitability sub-score of 1.3 and an impact sub-score of 6.0. The impact term is doing the work; the exploitability term is low because the attacker needs a victim to click.
Affected versions and the fix
The advisory lists four fix releases for three affected branches. Note that the newest branch got a minor bump, not a patch — if you are on 11.9 you cannot simply take 11.9.3.
| Branch | Affected | Fixed |
|---|---|---|
| 11.9.x | up to and including 11.9.1 | 11.9.2 |
| 11.8.x | up to and including 11.8.5 | 11.8.6 |
| 11.7.x | up to and including 11.7.10 | 11.7.11 |
| 11.10.x | — | 11.10.0 |
Mattermost credits z01dberg for the report under its responsible disclosure policy. The advisory page also notes that detailed write-ups appear 30 days after a fix ships, so expect a fuller technical account of the mechanism in late November.
Check your build before you plan the window
Version drift is the normal state of self-hosted software: the container tag says one thing, the running binary says another. Read the binary.
#!/usr/bin/env bash
# mattermost-cve-2026-93851-gate.sh
# Exit 1 if the running Mattermost build sits inside the affected ranges of
# CVE-2026-93851 (MMSA-2026-00742). Pass a version as $1 to test offline.
set -euo pipefail
ver="${1:-$(mattermost version 2>/dev/null | grep -oE '[0-9]+\.[0-9]+\.[0-9]+' | head -1)}"
if [ -z "${ver:-}" ]; then
echo "cannot read the Mattermost version" >&2
exit 2
fi
IFS=. read -r major minor patch <<< "$ver"
affected=0
if [ "$major" -eq 11 ]; then
if [ "$minor" -eq 9 ] && [ "$patch" -le 1 ]; then affected=1; fi
if [ "$minor" -eq 8 ] && [ "$patch" -le 5 ]; then affected=1; fi
if [ "$minor" -eq 7 ] && [ "$patch" -le 10 ]; then affected=1; fi
fi
if [ "$affected" -eq 1 ]; then
echo "AFFECTED $ver — upgrade to 11.10.0, 11.9.2, 11.8.6 or 11.7.11"
exit 1
fi
echo "ok $ver"
Run it against the affected and fixed builds to confirm it agrees with the table before you wire it into a check: ./mm-gate.sh 11.9.1 must exit 1, ./mm-gate.sh 11.9.2 must exit 0. The version is taken as an argument for exactly that reason — a gate you cannot test offline is a gate you cannot trust.
What to do beyond patching
Patching closes this specific hole. Three things worth doing while you are in there:
- Inventory your plugins. Every installed plugin is a server-side binary someone deployed. If nobody can name why each one is installed, that is your real finding, not the CVE.
- Treat the Agents plugin as a privileged surface. It renders model output and reacts to message content in an administrator's client. That is a new class of exposure for a chat server, and it deserves the same review you give a browser extension.
- Reconsider who can post where administrators read. The precondition here was mundane post permission in an administrator-visible channel. If your announcement or ops channels are open to everyone, an attacker does not need to compromise anything to reach step one.
The pattern to watch
The same advisory page, the same week, carries MMSA-2026-00723 / CVE-2026-116516 — a MEDIUM in the same Agents plugin, where summarising certain Zoom chat transcriptions could crash it because line length was not validated, fixed in the same four releases. Two advisories in one plugin inside a week is not a scandal; it is what a young, fast-moving surface looks like.
We mapped the same governance gap across the MCP ecosystem in 15,465 MCP Servers and Almost No Governance — there too the publication path has no gatekeeper. Agent plugins in chat platforms are where browser extensions were in 2010: enormously useful, installed casually, and holding privileges nobody audited. The lesson from this CVE is narrower than "AI is dangerous" and more useful than "patch your server". It is that the plugin decides what your chat server can do, and a chat message is now a delivery mechanism for requests your administrators did not mean to send.
What we do not know
No public exploitation is recorded in the CVE feeds as of 29 October 2026, and no proof-of-concept has been published. The advisory does not describe the exact request the administrator's browser is tricked into sending, only that the target post ID was unvalidated; the full mechanism is expected with the 30-day disclosure.
One data-quality note for anyone building a detector: the CVE feed's affected-product table lists mattermost_desktop, while the advisory text and the fix versions describe a server-side plugin issue affecting the Agents plugin. Treat the product mapping as unreliable and key your checks on the server version and the installed plugin set, which are the two things the advisory actually names.
author
Editor & Operator
MrDodgerX (MrDodgerX) is an enterprise geospatial systems engineer in Malaysia and the editor behind hejes.my. He runs the briefing pipeline, curates and edits every draft, verifies each factual claim against primary sources, and presses the publish button — nothing ships without his review. He writes about AI, security and Malaysia's tech scene from a builder's perspective. Professional profile and code: the links below.
sources
fact-check verified · 29 OCT 2026
Source details and access dates
- Mattermost Security Updates (MMSA-2026-00742 / CVE-2026-93851)Mattermost · accessed 29 OCT 2026
- CVE-2026-93851 — Path traversal in Mattermost Agents plugin mention reminderCVE Feed · accessed 29 OCT 2026
- Mattermost plugins — overviewMattermost Developers · accessed 29 OCT 2026
related

LMCache CVE-2026-105192: Unpatched 9.8 RCE
LMCache unpickles data straight off an unauthenticated socket. One message is RCE, your exposure is one flag wide, and no patch exists.

FortiMail Zero-Day: Disable IBE and Hunt
A 9.8 path traversal plus NULL byte flaw in FortiMail is under attack. The workaround is three CLI lines, and it does nothing for boxes already written to.

Hundreds of AI Agents Hit 440+ PaperCut Instances
GreyNoise tracked AI agents exploiting PaperCut flaws, hitting 440+ instances across 395 organizations in 48 countries. Campaign mechanics and patch guidance.
join the feed
one fresh insight per week. no spam, ever.
weekly digest. unsubscribe anytime.