The Security Tool Was the Backdoor: A Poisoned cPGuard Update Ran as Root on ~2,950 Hosting Servers

The Security Tool Was the Backdoor: A Poisoned cPGuard Update Ran as Root on ~2,950 Hosting Servers

The Security Tool Was the Backdoor: A Poisoned cPGuard Update Ran as Root on ~2,950 Hosting Servers

For about 70 minutes on November 2, the most dangerous software on thousands of Linux hosting servers was the security product installed to protect them. At 02:06 UTC, an attacker used compromised credentials to enter OPSSHIELD's update distribution infrastructure and push a malicious script disguised as cPGuard versions 5.88.12 through 5.88.19. Customer agents installed it between 02:08 and 03:20 UTC, and it ran as root — reading /etc/shadow, /etc/passwd, root's SSH private keys, the MySQL root password from /root/.my.cnf, and root's shell history, with a small number of servers also leaking cloud storage keys from /root/.s3cfg. OPSSHIELD's incident disclosure counts about 2,950 customer servers that ran the script, which sent everything it collected back to the attacker.

The intrusion path is a small-vendor supply-chain story in miniature, and the timeline is precise. On November 1 at 15:38 UTC the attacker began probing OPSSHIELD's billing portal, registered a customer account under a false identity at 16:34, and ordered a free cPGuard trial — installing it at 21:17 to study how the agent talks to the licensing system. cPGuard's own defenses flagged and banned the attacker twice. At 22:50 the attacker sent crafted requests to the billing portal's callback pages and by 22:58 had command execution on the billing and licensing server, using a stolen access key to upload the fake update at 02:04 on November 2. The attacker also copied three of OPSSHIELD's cloud servers and took billing records for about 7,900 customer accounts — names, companies, postal addresses, emails, and bcrypt-hashed passwords. The attack was detected within 12 hours of the poisoned versions going out.

OPSSHIELD's remediation guidance is unusually concrete. Version 5.88.26 removes the backdoor the script installed, and a follow-up update removes the scheduled task it left behind — the indicators to check for are /etc/cron.d/i93677, /var/tmp/.i93677/, and /var/tmp/.cpg91511_sh. Admins can search their update log with grep 'Successfully updated' /opt/cpguard/logs/install_log — any entry from 5.88.12 to 5.88.19 means the server pulled an affected update and should be treated as potentially compromised regardless. The company offers a lookup that checks server IPs and hostnames against its incident records, and it is shipping another precautionary update that removes its own support-access SSH keys from customer servers.

The credential reality is the hard part: passwords the script copied may still work, so every server that ran it needs root and account passwords reset, MySQL and database passwords changed, SSH keys replaced, and S3 or DigitalOcean tokens rotated — with any server reachable via SSH from an affected host treated as exposed until its keys are replaced too. If OPSSHIELD confirms an attacker logged into a server manually, that server should be rebuilt from a clean backup taken before November 1.

What Should You Do?

  1. Update and verify. Run cpgcli update --start on every cPGuard server and confirm the version is 5.88.26 or later; check the update log for any 5.88.12–5.88.19 entries.
  2. Check for the artifacts. Delete any of the three indicator files if present, and use OPSSHIELD's IP/hostname lookup — a missing record does not prove a server was unaffected.
  3. Rotate everything root-level. Assume /etc/shadow, root SSH keys, the MySQL root password, shell history, and any cloud keys are known to the attacker — reset passwords, replace keys, rotate tokens, and treat SSH-reachable neighbors as exposed.
  4. Reset your billing portal credentials too. All OPSSHIELD customers should change their manage.opsshield.com password and enable two-factor login.

The WAF Angle

cPGuard's own marketing calls it "ultimate web server security" — malware scanner, WAF, abuse-IP blocking. This incident is the supply-chain inversion of that promise: the WAF agent itself became the root-level payload delivery channel, because an update channel is functionally a remote-code-execution path that every customer grants you voluntarily. The defenses that would have limited the blast radius are the ones hosting operators should demand from every vendor: signed, staged updates with canary populations; runtime alerts when a "security agent" starts reading /etc/shadow and /root/.ssh; and secrets architecture that assumes root compromise (hashed passwords everywhere, no plaintext database credentials in dotfiles, no cloud keys lying in root's home). Note also what the attacker did first — they registered as a customer, took the free trial, and studied the agent-licensing protocol, the same trusted-interface reconnaissance pattern behind the trojanized Claude Code extension and the ConnectWise supply-chain compromise. Your security vendor's update pipeline is attack surface; treat it that way.

Sources