cPGuard is a security agent for Linux web-hosting servers: a malware scanner and firewall that runs alongside cPanel, DirectAdmin and Plesk, sold by OPSSHIELD and offered by hosts including Namecheap on their VPS and dedicated plans. Like most agents of its kind it runs as root and updates itself. Between 02:08 and 03:20 UTC on 2 November 2026, about 2,950 of those servers installed an update that was not OPSSHIELD's. It copied the servers' password hashes, SSH keys, database root passwords and shell history, left a root shell behind, and sent everything to the attacker.

No bug in cPGuard was exploited. The attacker signed up as a customer, used a free trial to learn how the agent talks to its vendor, broke into the vendor's billing and licensing server, found the key that publishes updates, and used it. The whole thing, from the first probe to the last infected server, took under 12 hours. OPSSHIELD's own disclosure gives the timeline to the minute, and it is one of the clearest public accounts yet of how a security vendor's update channel becomes the attack.

[INFO]
Primary sources: OPSSHIELD's two notices, the "cPGuard security advisory: Update channel compromise" (2 and 3 November) and "Security Incident Disclosure: manage.opsshield.com" (3 November), both read on 12 November. The server count quote is from BleepingComputer, 3 November. Where the two notices differ, the article gives both.
◈ interactive artifact
Seventy Minutes on the Update Channel
Scrub from the first probe on 1 November to the disclosure on 3 November. On one side, OPSSHIELD's servers as its own timeline describes them; on the other, 2,950 customer servers taking the update, and one hosting admin's terminal running the advisory's checks.

//The Timeline, in OPSSHIELD's Words

All times UTC, from the disclosure. On 1 November at 15:38 the attacker began probing manage.opsshield.com, the billing portal. At 16:34 it registered a customer account under a false identity and ordered a free cPGuard trial. At 21:17 it installed the trial on its own server to study how the agent talks to the licensing system; cPGuard itself flagged and banned it twice, at 21:40 and 21:51. At 22:50 it sent crafted requests to the billing portal's callback pages, and at 22:58 it could run commands on the billing and licensing server and began collecting data.

On 2 November between 02:04 and 02:09 it uploaded a fake cPGuard update to the update storage, using an access key it had taken. From 02:08 to 03:20 cPGuard agents installed it. At 02:13 the attacker also copied three of OPSSHIELD's cloud servers. OPSSHIELD revoked the stolen key and published a clean update, 5.88.26, the same day, and on 3 November traced the entry point and published the disclosure.

OPSSHIELD says it has ruled out SSH, the control panel, FTP and API tokens as the way in, and has patched two possible entry points in the billing application. Its logs do not record the exact request that gave the attacker command execution, so the specific vulnerability is not confirmed.

//What the Fake Update Did

The malicious releases carry the version numbers 5.88.12 to 5.88.19. Per the disclosure it was "a short script", in four variants named cpguard-v58812-i91511*.sh, running as root. It copied /etc/shadow and /etc/passwd, the contents of /root/.ssh, /root/.my.cnf where present, root's shell history and, on a few servers, /root/.s3cfg. It left a hidden setuid-root shell at /var/tmp/.cpg91511_sh, files under /var/tmp/.i93677/, and a scheduled task at /etc/cron.d/i93677.

That list is a hosting server's keys to everything else. Shadow hashes can be cracked offline. Root's SSH directory holds the private keys it uses to reach other machines and the list of keys allowed in. .my.cnf is the database root password in plain text. Shell history is where administrators paste tokens and passwords. And because the agent is installed on servers that host other people's websites, each of those servers is itself a supply chain.

OPSSHIELD's guidance is direct: for any server where cPGuard was online between 02:08 and 03:20 UTC on 2 November, "assume its root-level secrets are known to the attacker". Servers that were not online in that window did not receive the update.

//Two Notices That Do Not Quite Agree

The first advisory, written while the incident was still being worked, says the script uploaded what it collected "to OPSSHIELD's licensing server" under a compromised account, and gave a preliminary estimate of 32 collected files retrieved by the attacker, each one a customer server. BleepingComputer's report on 3 November follows that version. The later disclosure says the files went to "a server the attacker controlled" and does not repeat the 32 figure. The first advisory says 5.88.25 or later removes the backdoor; the disclosure names 5.88.26 as the clean update, with a follow-up to remove the scheduled task. Read together, the safe reading is the later one: assume the data left, update to the latest release, and check for the cron entry yourself.

On scale they agree. An OPSSHIELD spokesperson told BleepingComputer "We detected 2946 servers were affected"; the disclosure rounds it to about 2,950. Because agents check for updates on a schedule, not every online server necessarily took the update, which is why the advice is phrased around the time window rather than a list.

//The Vendor's Own Data

The billing server held more than the update key. The attacker copied about 7,900 billing-portal accounts: name, company, postal address, email and a bcrypt hash of the password. OPSSHIELD is resetting portal sessions, re-issuing every cPGuard licence key pair so the attacker's copies stop working, rotating its own SSH keys, cloud tokens and API keys, and rebuilding its billing, licensing and management servers from clean images. On its own servers it found a gsocket backdoor and Go implants that hide from standard tools, and recommends checking from a trusted boot environment.

//Why This Pattern Keeps Working

This is our reading, not OPSSHIELD's.

The update key lived next to the billing app. A customer-facing PHP application and the credential that ships root code to thousands of servers were reachable from the same machine. Whatever the entry bug turns out to be, that adjacency is what turned a billing-portal compromise into a supply-chain attack.
The trial was the reconnaissance. Anyone can sign up and get a working agent. The attacker used it to learn the licensing protocol from the inside, and the product's own detection banned it twice without stopping it. Free trials of security agents are a gift to anyone studying how to impersonate the vendor.
Auto-update as root, without a second check. The agents installed what the update storage offered. A signature check against a key that was not on the billing server, or a staged rollout that held a release for an hour, would each have broken this chain. Neither is exotic. This site's goxlr.live and DebugBear articles are the same lesson at the user's end: the delivery path was trusted, so what came down it was trusted.

//If You Run cPGuard

From OPSSHIELD's two notices: update with cpgcli update --start and confirm the version with cpgcli --version; check /opt/cpguard/logs/install_log for "Successfully updated" entries for 5.88.12 to 5.88.19; confirm /etc/cron.d/i93677, /var/tmp/.i93677/ and /var/tmp/.cpg91511_sh are gone (their absence does not prove the script never ran). Then change root and every local password, the MySQL root password and anything in .my.cnf or .s3cfg; generate new root SSH keys and remove the old ones from every authorized_keys that trusted them; rotate anything that appears in root's shell history; and treat servers reachable with those keys as exposed. If OPSSHIELD told you the attacker logged in to your server by hand, rebuild it from a backup taken before 1 November. Reset your billing-portal password and turn on two-factor login.

[IOC]
Vendor: OPSSHIELD (cPGuard); entry: billing and licensing server via manage.opsshield.com, from a customer account registered 1 Nov 16:34 UTC
Malicious releases: cPGuard 5.88.12 to 5.88.19, installed 02:08 to 03:20 UTC on 2 November 2026; about 2,950 servers (2,946 per OPSSHIELD to BleepingComputer)
Scripts: cpguard-v58812-i91511*.sh (four variants; SHA-256 values in OPSSHIELD's disclosure)
On customer servers: /var/tmp/.cpg91511_sh (setuid-root shell), /var/tmp/.i93677/, /etc/cron.d/i93677; first advisory also lists /var/tmp/.cpg91511/loot.tgz
Collected: /etc/shadow, /etc/passwd, /root/.ssh, /root/.my.cnf, root shell history, /root/.s3cfg on a few servers
Attacker address: 94.183.188.91 (sometimes with a fake cPGuard/5.88.11 user agent); web-root files i90248_probe.txt, i90573_probe.txt, i90573_sh.php, i90573_x.php; receivers matching *_rx.php
Vendor data: about 7,900 billing accounts (name, company, address, email, bcrypt hash)
Clean releases: 5.88.25 or later (first advisory), 5.88.26 (disclosure) plus a follow-up removing the cron entry

//Sourcing and Uncertainty

The timeline, entry path, collected files, artefacts, indicators, the 7,900 accounts, OPSSHIELD's actions and the "assume its root-level secrets are known" guidance are from OPSSHIELD's disclosure. The CDN wording, the 32-file estimate, the licensing-server upload, the 5.88.25 version and the install-log check are from its first advisory. The spokesperson's 2,946 is from BleepingComputer. The product description is from BleepingComputer and OPSSHIELD. The section on why the pattern works is our analysis.

What is unknown. The exact vulnerability in the billing application. Whether the stolen access key was also a signing key, and whether agents verified any signature at all. What the three copied cloud servers contained. How many customer servers the attacker logged in to by hand. Who the attacker is: OPSSHIELD does not say, and neither does this article.