The espionage campaign where the hackers forgot to delete the chatbot's comments
Palo Alto's Unit 42 says a seven-month intrusion into government Linux servers used scripts that still carried AI assistant chatter. Microsoft, separately, says North Korean crews are now routinely automating their day jobs with the same tools.
Key points
- Palo Alto Networks' Unit 42 says attackers hit government and state-linked Linux servers from October 2025 to April 2026, leaving AI-generated code with the chatbot's own commentary still inside it.
- The campaign, tracked as CL-UNK-1211, tried to steal government mailboxes and dump sensitive databases.
- Intruders planted fake red-team markers so defenders would mistake the break-in for a hired security test.
- Microsoft reported separately that North Korean groups including Jade Sleet, Onyx Sleet, Coral Sleet, Diamond Sleet and Pearl Sleet are now using AI to generate lures and tooling at speed.
- Onyx Sleet and Pearl Sleet have adopted ARTEX, an open-source AI hacking framework, for live operations.
For seven months, somebody quietly rummaged through a set of government Linux servers. Linux is the free operating system that runs most of the internet's back rooms, including a lot of government infrastructure.
The interesting bit isn't that they got in. It's what they left behind.
Researchers at Palo Alto Networks' Unit 42 found that the scripts dropped on those servers still contained the little conversational notes that large language models, the technology behind tools like ChatGPT, tend to sprinkle through generated code. Things like "here is the reasoning" and polite developer asides. The attackers either didn't read what their assistant produced, or didn't care.
Unit 42 is tracking the activity as CL-UNK-1211, where "CL" refers to cluster and "UNK" stands for unknown. Targets were government bodies and state-affiliated organisations across a window of October 2025 to April 2026.
What were the attackers trying to do?
Steal information, mostly. Unit 42 says the intruders tried to dump sensitive databases and pull down government email mailboxes, then use those mailboxes to get into social media accounts linked to officials. Classic intelligence work, done with modern plumbing.
The focus was on Linux boxes because that's what the target networks ran. Custom backdoors, which are hidden entry points coded to work on more than one operating system, were mixed in with free hacking tools pulled from the internet. Web shells, small malicious files that let an attacker control a web server through a browser, were also sprayed across the environments they reached.
How did they try to hide?
By pretending to be the good guys. Along with the usual tricks, switching off the server's antivirus and wiping system logs, the attackers planted markers that mimicked a hired penetration test, the kind of friendly break-in companies pay security firms to run. If a defender stumbled on the activity, the hope was they'd shrug and assume it was authorised.
First reported by The Hacker News, the campaign is one of the clearer public examples of attackers using AI to write their malware rather than just their phishing emails.
Is this connected to North Korea's AI push?
Unit 42 hasn't named a country. But we covered Microsoft's own account of North Korean AI adoption the day before this report landed, and the picture rhymes.
Microsoft says the group it calls Jade Sleet stood up three fake companies in a single week. Onyx Sleet produced dozens of tailored phishing lures in minutes. Coral Sleet built a new tool for poisoning private source code repositories, meaning the places where companies store the master copies of their software. Diamond Sleet used AI to invent believable personas and front companies. Onyx Sleet and Pearl Sleet have both picked up ARTEX, a free AI-driven hacking framework, for real targets.
| Group | Reported AI use |
|---|---|
| Jade Sleet | Three front companies spun up in one week |
| Onyx Sleet | Dozens of tailored lures generated in minutes; uses ARTEX |
| Coral Sleet | Built a tool to trojanize private code repositories |
| Diamond Sleet | Persona and front company creation |
| Pearl Sleet | Uses ARTEX for operations |
My read: the CL-UNK-1211 scripts, with their leftover AI chatter, are the clumsy end of a trend Microsoft is describing at the professional end. Attackers are starting to treat an LLM the way the rest of us treat a junior colleague who works fast. Sometimes you check the work. Other times it ships as-is.
For ordinary readers there's nothing to patch here. The useful point is narrower: when a government mailbox gets emptied, expect the social accounts tied to it to be the next target, and treat password reset emails from officials' accounts with suspicion for a while afterwards.
