Symantec: AI-written scripts turned a clumsy retail intrusion into fast lateral movement via SolarWinds Orion
Symantec says attackers at a US retailer switched from typo-filled commands to AI-written scripts and moved through the network in minutes. No new tricks — just faster.
Symantec says an intruder inside a US retailer's network went from typo-ridden, hand-typed commands to polished, AI-written scripts overnight, and then reached a second server within about ten minutes of looking inside the company's SolarWinds Orion monitoring system. No data theft or encryption was seen, but the case shows what AI help looks like in a real intrusion: the same old techniques, done faster and with fewer mistakes. If you run Windows servers, SQL Server or a network-monitoring platform, check the hardening points and published indicators below.
What happened, in plain English
On November 12, 2026, Symantec's Threat Hunter Team (part of Broadcom) published an analysis of a September 2026 intrusion at an unnamed US retail company.
The attackers first showed up on Microsoft SQL Server database machines on September 18, apparently running commands through the database service itself. How they got in originally is unknown. On September 22, working on a Windows web server, they were clumsy: typos such as "pwoershell", and the same blocked tool re-run about five times in 20 minutes.
The next day, their behaviour changed. Instead of short commands, they ran about 50 long, tidy PowerShell scripts (PowerShell is the scripting language built into Windows) in roughly two hours. The scripts had labelled output, error handling and comments that reasoned about earlier results — the hallmarks of text written by a large language model (LLM), the technology behind chatbots such as ChatGPT. Symantec says it assesses the AI assistance "with high confidence", but could not tell which model was used.
With that help, the attackers collected stored passwords and found the company's SolarWinds Orion server — a tool IT teams use to watch every device on the network. That gave them a map of the network and a way onto another server about ten minutes later.
Think of a burglar who fumbles with the first lock, then calls an expert on the phone who talks them through the rest of the house. The tools are the same; the burglar is just much faster.
Are you affected? What to do now
Symantec has not named the victim or the attackers, and there is no new vulnerability or patch here. Most readers have no emergency action — but the techniques are common, so this is a good moment to check the basics, especially if you run SQL Server, IIS web servers or an Orion-style monitoring platform.
Check exposure
- Search your firewall, proxy and DNS logs for the network indicators Symantec published, for example
204[.]194[.]49[.]129,166[.]0[.]25[.]106,154[.]81[.]37[.]136,155[.]254[.]22[.]215and108[.]165[.]20[.]205. The full list, including about 30 file hashes (SHA-256 fingerprints), is in the Symantec report. - Look for file names Symantec lists in unusual places, such as
bdreinit.exewithlog.dll,ask.exewithask.dll,server.ps1, and web shells namedfilemanager.asporqwdajhsd.asp. - Review whether the SQL Server feature that lets the database run operating-system commands (
xp_cmdshell) is enabled anywhere. Symantec says the earliest activity "likely" used it.
Harden (our recommendations, based on the techniques in the report)
- Treat monitoring servers as crown jewels. A platform like SolarWinds Orion holds a list of every device and often the credentials to reach them. Restrict who and what can log in to it, and keep it off the general server network where possible.
- Lock down SQL Server. Keep
xp_cmdshelldisabled unless there is a documented need, run the database service under a low-privilege account, and use strong, unique passwords for built-in accounts such assa— the attackers tried to guess them. - Remove stored passwords from configuration files and file shares. The attackers searched point-of-sale sync folders and saved logins from tools such as FileZilla, WinSCP, PuTTY and Remote Desktop.
- Use unique local administrator passwords (for example with Microsoft LAPS, the Local Administrator Password Solution) and protect who can read them — the attackers went looking for LAPS passwords on what appeared to be a domain controller.
- Watch for unusual parent processes. Long PowerShell scripts started by the web server process (
w3wp.exe) or by SQL Server are a strong warning sign, as are tunnelling tools such as Chisel. - Plan for speed. If an attacker can move in minutes, your alert-to-containment process needs to be measured in minutes too. Make sure someone can isolate a server out of hours.
The expert view
The interesting part of this report is not any single technique — DLL sideloading, a potato-family privilege escalation, credential dumping from registry hives, password spraying and tunnelling are all well documented. Symantec says so directly: "AI supplied no new techniques or tactics, only a new way of meeting objectives." What changed, in its words, is "the speed and reliability with which the attackers applied them."
The evidence for AI is behavioural, and it is worth understanding because defenders can look for it too. Median command length jumped from about 45 characters to more than 500; typos disappeared; scripts carried structured banners, capped output and comments that stated a hypothesis, cited previous results and revised conclusions. Humans pasting one-shot commands into a web shell rarely write like that.
Symantec also separates "AI-assisted" from "autonomous". The gaps between commands never dropped below seven seconds, with pauses of two to eight minutes before changes of strategy, and there was no machine-gun scanning or fixed-interval beaconing. Symantec judges this "more consistent with an operator relaying output to an AI tool than with an autonomous agent" — a human in the loop, copying results into a chatbot and pasting back its suggestions. It rates that operator-in-the-loop model with moderate confidence, and notes a change of comment style that may mean more than one model was used.
That distinction matters. Much recent reporting, including the Google Threat Intelligence Group's November tracker and multiple ARTEX reports, focuses on fully automated agents. This case is the more mundane — and probably more common — pattern: a mid-skill operator using an assistant as a senior colleague. Symantec says this was "one of dozens of AI-assisted attacks" its team has investigated in recent months, with other cases at a Vietnamese university, an East African technology provider and a South Asian hydroelectric company.
The Orion pivot is the textbook illustration. The attackers queried Orion's local database for its list of monitored nodes, then abused the RabbitMQ message broker that ships with Orion — specifically its Erlang cookie, a shared secret that lets cluster nodes trust each other — to run commands on a second server. None of that is novel, but it requires knowing an obscure product's internals. An LLM collapses that research time from hours to minutes, which is exactly the gap that turns a stalled intrusion into lateral movement.
On attribution, Symantec is careful. The motive looks financial: point-of-sale credential searches and one address linked to a card-theft operation. A staging server overlaps with a September 22 Gambit Security report on an operator using open-source AI tools against online retailers, and Symantec says the operations "may be linked". The loader pairing has been associated with ShadowPad, and the Yuze tunnelling tool with a China-nexus ransomware group, but Symantec stresses those tools are shared and the links are not conclusive. We do not attribute this intrusion to any group or country.
What is still unknown: the initial access vector, which AI model or service was used, and whether the provider detected the abuse. For defenders, the takeaway is uncomfortable but practical: AI does not need new exploits to hurt you. It makes every forgotten credential, over-privileged service account and flat network segment cheaper to find and use — so the fundamentals now have to hold up against a faster attacker.
Official sources
- Symantec Threat Hunter Team: AI-assisted attack against a US retailer (November 12, 2026) — primary report, including the full IOC list
- Gambit Security: autonomous AI agents against online retailers (September 22, 2026) — earlier report cited by Symantec as possibly linked
- Coverage: Check Point Research weekly threat intelligence report, November 16, 2026
- Coverage: Dark Reading — Inside an AI Attack Against a US Retailer