Skip to the content
AI and Cybersecurity Learning and Practice Center Test automation Open the Learning Center

IncidentAI-enabled attacks

CL-UNK-1211: AI-written scripts, a Go backdoor and a fake sudo prompt in a six-month espionage campaign against government Linux servers

HighPalo Alto Networks Unit 42

Palo Alto Networks Unit 42 published on October 15, 2026 its account of CL-UNK-1211, a campaign against government entities and state-affiliated infrastructure that ran from October 2025 to April 2026 on Linux and database servers; it names no countries. The attackers' PHP and shell scripts carried conversational annotations and step-by-step reasoning of the kind a language model writes, with section headers, heavy Base64 encoding and rapid switching between PHP, JavaScript, Bash and Python, which Unit 42 reads as code generated by a model or run by an automated agent, without proof of autonomous operation; one mailbox-archiving script had broken variable syntax and would have failed. They kept access with web shells, including one disguised as a Plesk driver, and by setting auto_prepend_file in every PHP configuration so their code ran before any script; they deployed SysGoRAT, a Go remote access trojan for Linux and Windows that talks over WebSockets and persists through systemd, cron, the Windows registry and scheduled tasks; and they replaced .bashrc with a sudo alias that mimics the password prompt and logs the password to a file named after SolarWinds, which was not compromised. They harvested administrator logins from modified login pages, searched configuration and history files for mail and database passwords, read /etc/shadow, enumerated mailboxes for social media password-reset emails, dumped database tables to CSV and sent them out with netcat and to public file hosts. They cleared logs, backdated files, deleted single-use scripts, planted a "RedTeamOps2025" key and a "pentest123" password to look like an authorized test, and tried to kill Cortex XDR, which was blocked. Privilege escalation used LinPEAS and PwnKit (CVE-2021-4034); reconnaissance included a custom scanner for React2Shell (CVE-2025-55182). Unit 42 has not determined whether the actor is a state or criminal one.

Replay the attack step by step in the Learning Center

How it happened

  1. Scripts with the annotations and reasoning of model-written code, in four languages, deployed against PHP and Plesk servers.
  2. Persistence through every PHP configuration's auto_prepend_file, a Go backdoor and a sudo alias that captures passwords.
  3. Credentials, mailboxes and database tables collected and sent to public file-hosting services; logs cleared and red-team decoys planted.

Impact

Administrator credentials, mailboxes and database contents of unnamed government entities over six months; exfiltration was attempted, and Unit 42 does not confirm how much left.

The fix

Restrict outbound traffic from servers, alert when standard utilities spawn interactive shells, watch for log clearing, timestamp changes and attempts to stop endpoint protection, and treat red-team markers you did not plant as an intrusion.

Lessons

  • Model-written tooling is now ordinary in intrusions, and it is identifiable by its comments; it does not have to be good code to work.
  • A sudo alias in .bashrc defeats every password policy; the shell's startup files are part of the attack surface.
  • Attackers plant red-team markers so responders stand down; confirm any 'authorized test' with the people who would have authorized it.

Sources