
Maintenance script for UME macOS and Linux computers.
Two scripts with the same items and results: device-tools.sh for macOS and Linux, and
device-tools.ps1 for Windows. They do two things:
| # | Item | Systems | What it does |
|---|---|---|---|
| 1 | Claude Code extension check | macOS, Linux, Windows | Looks for the traces left by the backdoored Claude Code extension (versions 2.1.312 and 2.1.313 from the Cursor extension marketplace). Read-only: it reports, it changes nothing. |
| 2 | Ubuntu network protection for CVE-2026-98432 | Ubuntu | Applies Ubuntu's published mitigation for a critical Linux kernel flaw (remote code execution). |
These commands always run the latest release.
macOS and Linux (Terminal). Needs sudo (your computer password):
f=$(mktemp) && curl -fsSL https://github.com/ScudraServicos/ume-device-tools/releases/latest/download/device-tools.sh -o "$f" && sudo bash "$f"
Add --status (check only, changes nothing) or --remove (undo the Ubuntu protection) at
the end of the same command:
f=$(mktemp) && curl -fsSL https://github.com/ScudraServicos/ume-device-tools/releases/latest/download/device-tools.sh -o "$f" && sudo bash "$f" --status
Without sudo (macOS and non-Ubuntu Linux only), the check covers the current user only.
Windows (PowerShell). Open the Start menu, type PowerShell, choose Run as
administrator, then paste:
[Net.ServicePointManager]::SecurityProtocol='Tls12'; $f="$env:TEMP\device-tools.ps1"; iwr -UseBasicParsing https://github.com/ScudraServicos/ume-device-tools/releases/latest/download/device-tools.ps1 -OutFile $f; if ($?) { powershell -NoProfile -ExecutionPolicy Bypass -File $f }
It only checks; nothing is changed on Windows. Without administrator rights it checks the current
user only. -ExecutionPolicy Bypass applies to that one run, not to the computer's settings. The
first part turns on TLS 1.2 for that window, which older Windows PowerShell needs to download
from GitHub.
To run one exact version, use releases/download/<tag>/device-tools.sh (or .ps1) instead of
releases/latest/download/... (tags are listed under Releases).
The two items run in separate sections, each ending with its own Result: line, and a
SUMMARY at the end repeats both:
| Item | Result | What to do |
|---|---|---|
| 1. Claude Code extension check | OK | Nothing. No traces found. |
ATTENTION | Do not delete anything. Contact the Security team now. | |
| 2. Ubuntu network protection | ACTIVE | Nothing. When Ubuntu releases the fix: update, restart, run again with --remove. |
NOT ACTIVE (--status) | Run the command without --status to turn it on. | |
NOT NEEDED | Nothing. The computer does not run Ubuntu (always the case on Windows). | |
REMOVED (--remove) | Nothing, if the official fix is installed. |
| Exit code | Meaning |
|---|---|
| 0 | OK |
| 1 | Error, or (--status) the Ubuntu protection is not active |
| 2 | Wrong arguments |
| 3 | Traces of the backdoored extension found: contact the Security team, delete nothing |
Source: Cursor security notice and Phorion's analysis.
On October 29, 2026, modified builds of the Claude Code extension (2.1.312 and 2.1.313) were served by the Cursor extension marketplace. When the editor loads them, they add an attacker SSH key, install a startup item, open a reverse SSH tunnel and send host details to the attacker. Updating or removing the extension does not remove these traces.
The check runs for every local account when run with sudo (macOS, Linux) or as administrator
(Windows). It reports:
| Trace | macOS and Linux | Windows |
|---|---|---|
Backdoored extension file (contains 134.209.82.225) | ~/.<editor>/extensions/anthropic.claude-code-*/extension.js | %USERPROFILE%\.<editor>\extensions\anthropic.claude-code-*\extension.js |
Attacker SSH key (svc-telemetry, SHA256:0Z6R+DF64Do2nwQic8a7kkfDLNNk4IiK+2U2Rci4htk) | ~/.ssh/authorized_keys | %USERPROFILE%\.ssh\authorized_keys, %ProgramData%\ssh\administrators_authorized_keys |
| Startup script and tunnel key | ~/.config/.smt, ~/.config/.smtk | %USERPROFILE%\.config\smt.cmd, %USERPROFILE%\.config\smtk |
| Startup item | ~/Library/LaunchAgents/com.apple.sysmond.plist (macOS), @reboot crontab entry (Linux) | scheduled task SysMon (or any task running smt.cmd), sysmon.cmd in the Startup folder |
| Tunnel key in the temporary folder | /tmp/.stk_*, $TMPDIR/.stk_* | %TEMP%\.stk_* |
| Running tunnel or retry loop | ssh/autossh to 134.209.82.225 or with .stk_/.smtk; shell running .config/.smt | ssh.exe to 134.209.82.225 or with .stk_/smtk; cmd.exe running smt.cmd or sysmon.cmd |
| Open connection to the attacker server | 134.209.82.225 | 134.209.82.225 |
Having version 2.1.312 or 2.1.313 installed is only noted, not counted: the backdoored builds came from the Cursor marketplace, and Cursor reports no other distribution as affected, so the file content decides.
On Windows the backdoor also tries to install and start OpenSSH Server, so the script notes when
the sshd service exists (not counted: many people install it on purpose). It also notes when
WSL is installed: with Cursor or VS Code connected to WSL, the extension can run inside WSL as
Linux, so run the macOS/Linux command inside WSL too.
If anything is found, do not delete it. The Security team handles removal and credential rotation, and needs the traces to check whether anyone logged in through the tunnel.
The real fix is an updated kernel. This protection reduces the risk until the fixed kernel is installed and running. It does not fully prevent exploitation.
| Step | Detail |
|---|---|
| Checks | Prints the running kernel and whether its changelog lists the CVE. |
| Installs | nftables, if missing (the only step that needs network access). |
| Rule | Writes /etc/cve-2026-98432.nft and loads table inet cve_2026_98432: ip protocol tcp ip hdrlength > 5 counter drop on the input hook, priority -310. Own table: existing firewall rules are untouched. Re-running never duplicates the rule. |
| Reboot | Enables cve-2026-98432-mitigation.service, a oneshot unit that reloads the rule at boot. |
/etc/nftables.conf | Only if nftables.service is already enabled: adds include "/etc/cve-2026-98432.nft", so systemctl reload/restart nftables keeps the rule. |
--remove deletes the table, the rule file, the unit and the include line. It leaves the
nftables package installed. Hosts protected with the earlier cve-2026-98432-mitigate.sh
(v2026.11.04) use the same names, so this script manages them too.
Impact: blocks incoming IPv4 TCP packets that carry IP options. Normal IPv4 TCP, TCP options, IPv6 and existing firewall rules keep working. TCP connections that depend on IPv4 options (rare, e.g. CIPSO-labeled networks) break. Containers with their own network namespace need the script run inside them too.
Differences from the advisory steps. The rule is the advisory's rule, unchanged. Boot
persistence uses its own systemd unit instead of enabling nftables.service, because that service
runs flush ruleset when it starts and would wipe rules managed by UFW, firewalld, Docker or
Kubernetes. The rule file starts with table + delete table, so reloading it never adds a second
copy. On Ubuntu 18.04 the nftables package enables nftables.service on install, so there the
include line is added too.