Skip to content
AI Security Wire

Published

- 10 min read

By

Symantec: AI Revived a Stalled Intrusion at a US Retailer

img of Symantec: AI Revived a Stalled Intrusion at a US Retailer

Incident Classification: Confirmed | Incident Type: Reported | Severity: Medium (our assessment) | Sector: Retail | Date of Activity: 18 to 29 September 2026 (latest dated activity in the report) | Documented by Symantec: 12 November 2026

Symantec’s Threat Hunter team documented a September 2026 intrusion at an unnamed US retail organization in which the attackers, blocked while working by hand, switched on the web server to what looks like AI-generated PowerShell, and moved to a second server about 10 minutes after first querying SolarWinds Orion. All incident facts in this article come from Symantec’s report; the victim is not named, and we found no evidence as of 14 November in Symantec’s post or Dark Reading of independent confirmation (Dark Reading derives from Symantec). Symantec did not identify the initial access vector or the AI model, and no data theft or encryption was observed.

Incident Summary

FieldDetail
VictimUnnamed US retail organization
Initial accessUnknown; Symantec did not identify the vector (Dark Reading reports no vulnerability was exploited, which Symantec’s post does not state)
Systems reachedInternet-facing web server and at least six further servers, most running Microsoft SQL Server; one appears to be a domain controller
AI evidenceLong, error-free labelled PowerShell scripts, and comments that read as a model’s reasoning
Human involvementOperator relaying output to an AI tool, moderate confidence; AI assistance itself, high confidence
Lateral movementOrion’s bundled RabbitMQ and its Erlang cookie, used to run commands on a second web server
Payload and toolingDLL sideloading loaders, ASP web shells, EfsPotato, Chisel, a process-tree-breaking tool “BreakChain”, a password sprayer
Observed impactCredential harvesting and lateral movement; no data theft or encryption observed
Likely motiveFinancial, per Symantec (point-of-sale password searches, infrastructure overlap with a card-theft operation)
AttributionNone; the AI model or service was not identified

Timeline

Dates come from Symantec’s report. The report describes activity on several computers; this is a condensed order.

  • 18 September: earliest malicious activity, on three SQL Server hosts. Commands run as cmd.exe under sqlservr.exe, most likely through xp_cmdshell. Labelled multi-part PowerShell scripts appear already on these hosts.
  • 21 September: activity on a Windows web server (commands under w3wp.exe): a short discovery burst, then a gap of about 12 hours.
  • 22 September: hand-typed commands, with typing errors such as “systemifo” and “pwoershell”. The attackers download a Bitdefender executable with a malicious DLL and text file, and security software terminates it roughly five times in under 20 minutes. They fetch a PowerShell reverse shell and write two ASP web shells.
  • 23 September: the attackers return, stage a second sideloading set (a G DATA signed executable), and are blocked again. They fetch the BreakChain utility. Then, for nearly two hours, they submit about 50 long, error-free scripts. They harvest credentials from configuration files, query Orion’s database, read the RabbitMQ Erlang cookie and run a command on a second web server about 10 minutes after first querying Orion. A second BreakChain build arrives about 75 minutes after the first was blocked.
  • 24 September: a fifth computer saves its SAM and SECURITY hives.
  • 25 September: larger-scale activity. A Chisel tunnel is detected, a new staging server is used, and a probable domain controller is searched for NTDS.dit, Group Policy Preferences passwords and LAPS passwords.
  • 26 to 29 September: a password sprayer runs on a seventh computer, continuing until at least 29 September.

Phase One: Hand-Typed Commands

Symantec describes the first activity on the web server as manual. Across 110 distinct actions there were three typing errors, each corrected within seconds, and command lengths averaged about 45 characters. The attackers re-ran the same blocked loader five times in about 20 minutes with no change of approach. The loader was a legitimate Bitdefender executable (bdreinit.exe) paired with a malicious log.dll and a payload in log.txt, a pairing Symantec notes has been documented with ShadowPad loaders but has also been used by several actors, so it does not identify the actor on its own.

Phase Two: The Switch

On 23 September the character of the activity changed. The median script was more than 500 characters, over ten times the length of the hand-typed commands, with no typing errors. Each submission bundled several probes, labelled each result with a banner such as ===FILES MENTIONING CONN STRINGS===, wrapped cmdlets in error handling and capped the output returned. Symantec says that layout suits a consumer reading results back through a narrow text channel and is common in LLM-generated code.

The strongest evidence was in the comments. Several scripts opened with text that stated a hypothesis, cited earlier results and revised a previous conclusion before acting. One recorded that earlier negative results about a server came from a SOCKS proxy channel known to report false successes, so the conclusion had to be re-measured from a host with a direct route. Symantec notes human operators do not usually write that kind of deliberation into a one-shot web shell command.

The gap between submissions never fell below seven seconds, and pauses of two to eight minutes came before changes of strategy. There was no scanner-style burst activity and no fixed-interval beaconing. Symantec reads this as an operator copying output into an AI tool and pasting the next command into the web shell. Similar labelled scripts appeared on 18 September on three database servers, and comments disappeared between 19 and 22 September and returned in a different style, which Symantec says may mean more than one model was used.

How the Lateral Movement Worked

The report describes this step in detail. The attackers collected credentials from the web server: connection strings, the FileZilla Server user database, FileZilla, WinSCP and PuTTY client settings, saved RDP files, IIS configuration and Windows logon registry values. They then located the SolarWinds Orion installation, queried its local SQL Server instance for the table of monitored nodes, and read the Zabbix agent’s pre-shared key.

They identified the RabbitMQ message broker that ships with Orion and read its Erlang cookie file (.erlang.cookie), the shared secret Erlang nodes use to authenticate to one another. They ran the erl.exe bundled with Orion, with the cookie, to call the RabbitMQ node on a second web server and run an operating system command. About 10 minutes had passed since they first queried the Orion service. The command was encoded as decimal character codes (the first decodes to whoami); later calls queried the registry for the SAM database, cached logons and LSA secrets.

Symantec’s own conclusion is that this is not a new technique: Erlang remote procedure calls, cookie theft and credential harvesting from configuration files are long documented. What AI assistance appears to have changed is the speed and reliability with which known techniques were applied: the same operators with the same access had achieved little the day before.

Tooling and Infrastructure

  • Sideloading sets: bdreinit.exe with log.dll and log.txt, then a G DATA signed ask.exe with ask.dll and ask.txt.
  • BreakChain: relaunches a target so it no longer appears as a child of w3wp.exe; the attackers used its WMI and shell modes.
  • Privilege escalation: EfsPotato, built on the server from base64 text decoded with Certutil and compiled with csc.exe; registry hives were then saved.
  • Persistence and tunnelling: two ASP web shells in the web application’s images folder, a Chisel tunnel, and an AES-encrypted PowerShell reverse shell.
  • Overlap: Symantec found shared loader files and a shared staging server with an August 2026 intrusion at a different organization, and similar loaders in a July 2026 intrusion at a securities firm. Neither showed signs of AI assistance. Symantec also cites a 22 September report from Gambit Security about a financially motivated operator using AI tools against online retailers, and says the shared address, retail focus and point-of-sale password searches suggest the operations may be linked, though the Gambit report does not mention the loaders described here. We did not review the Gambit report ourselves.

Indicators (defanged, from Symantec)

Network indicators Symantec lists for this intrusion include: 204[.]194[.]49[.]129 (payload staging, port 8080), 166[.]0[.]25[.]106 (reverse shell host and C2, port 443), 154[.]81[.]37[.]136 (script host), 155[.]254[.]22[.]215 (egress test target, named as an AI console in public reporting) and 108[.]165[.]20[.]205 (later staging server, port 8080). File indicators, including SHA-256 hashes for the loaders, server.ps1 and BreakChain, are in Symantec’s report.

Other AI-Assisted Cases in the Same Report

Symantec says this is one of dozens of AI-assisted attacks it has investigated since May 2026, with most in September, and briefly describes several others. These are Symantec’s summaries, and several cases remain under investigation:

  • A June 2026 web shell compromise at a university in Vietnam, with a script containing numbered notes in Chinese addressed to the operator.
  • An August 2026 intrusion at an East African ICT provider, about 61 hours inside the network, with 212 long scripts, and a search for mobile-money terms such as M-Pesa.
  • An August 2026 intrusion at a government hydroelectric company in a south Asian country, where one script still contained the placeholder YOUR_IP.
  • A July 2026 shared web hosting server in a south Asian country with about 970 commands about one second apart.
  • Three linked cases in south Asia sharing tools or infrastructure.

Symantec reports that in all cases AI supplied no new techniques, only a faster way of meeting objectives, and that after blocks attackers renamed or rebuilt tools within seconds or minutes.

Remediation and Response

Symantec’s report points readers to its protection bulletin and says that if an IOC is malicious and the file is available to Symantec, its endpoint products will detect and block that file. For response guidance, Symantec analyst Dick O’Brien told Dark Reading that “the retry pattern is a signal itself,” that blocking on behavior rather than file names or exact commands matters more, and that treating a blocked attempt followed by a variant as a high-priority alert is worth considering. He also said defenders should look at mean time to mitigate rather than mean time to patch.

Our detection suggestions, derived from the behaviors Symantec describes, not from vendor-published rules:

  • cmd.exe or PowerShell launched under sqlservr.exe or w3wp.exe, especially followed by credential file searches.
  • Reads of .erlang.cookie by anything other than the broker, and erl.exe run with a -setcookie argument from a web server context. This will also fire on legitimate Erlang and RabbitMQ processes on Orion hosts, so baseline it first.
  • csc.exe compiling source written under ProgramData, or Certutil decoding files there.
  • A blocked execution followed within minutes by a renamed, rebuilt or relaunched variant.
  • Saves of the SAM, SYSTEM and SECURITY hives.

What This Means

Our assessment: the useful lesson is the one Symantec states. The report documents no new technique, and it does not say how the attackers got in, so we cannot say what patching would have changed. Symantec’s O’Brien makes the point that patching is still important but cannot be the only countermeasure. What the report describes is a failed attack becoming a successful one because the operator found a workaround faster. That puts the weight on detection of the retry-and-vary pattern and on reducing what a compromised web server can reach, such as monitoring software that ships a message broker with a reusable shared secret. As Symantec cautions, the operator-in-the-loop reading is a moderate-confidence judgment, and the report cannot show how much of the other cases was automated.

References

Frequently Asked Questions

What did Symantec report?
Symantec's Threat Hunter team published a report, covered by Dark Reading on 12 November 2026, saying attackers broke into a US retail organization in September 2026, were repeatedly blocked while typing commands by hand on a web server, and the next day began submitting there long, error-free PowerShell scripts that appear to be AI-generated. About 10 minutes after first querying the victim's SolarWinds Orion software, they had used a bundled component to run commands on a second server. Symantec did not identify the initial access vector. Dark Reading reports that no vulnerability was exploited, but Symantec's post does not say so.
Was it an autonomous AI agent?
Symantec assesses AI assistance with high confidence but the operator-in-the-loop model with only moderate confidence. The gap between submissions never fell below seven seconds, with longer pauses before strategy changes, which Symantec says fits an operator relaying output to an AI tool better than an autonomous agent. The model or service involved could not be identified.
What can defenders take from it?
Symantec's analyst Dick O'Brien told Dark Reading that the retry pattern is itself a signal, that blocking on behavior rather than file names or exact commands matters more, and that treating a blocked attempt followed by a variant as a high-priority alert is worth considering. He also argued defenders should track mean time to mitigate rather than mean time to patch. Any specific detection logic beyond that, such as watching for RabbitMQ cookie reads, is our suggestion.