# Symantec - AI-assisted commands revived a stalled intrusion at a US retailer

- ID: symantec-ai-revived-retailer-intrusion-2026-11
- Reported: 2026-11-16
- Researcher: Symantec Threat Hunter Team
- Organization: Broadcom
- AI role: attacker_uses_ai_as_operator (AI as operator)
- Confidence in AI involvement: not assessed
- Research / borderline: false
- Last reviewed: 2026-11-17
- Entry page: https://openattacks.ai/attacks/symantec-ai-revived-retailer-intrusion-2026-11/
- Entry JSON: https://openattacks.ai/api/attacks/symantec-ai-revived-retailer-intrusion-2026-11.json
- Browse title: Symantec - AI-assisted commands revived a stalled intrusion at a US retailer
- Editorial summary: Symantec Threat Hunter Team (Broadcom; SECURITY.COM article created Nov 16, 2026) describes a September 2026 intrusion at a U.S. retailer. The initial access vector is unknown. The earliest activity, Sep…
- Evidence class (catalog annotation): Not classified
- Source-stated target environments: Not specified
- Stage hunts and conditional defenses: https://openattacks.ai/api/defenses/symantec-ai-revived-retailer-intrusion-2026-11.json

## Technique
Symantec Threat Hunter Team (Broadcom; SECURITY.COM article created Nov 16, 2026) describes a September 2026 intrusion at a U.S. retailer. The initial access vector is unknown. The earliest activity, Sep 18, was on three Microsoft SQL Server hosts, with cmd.exe running under sqlservr.exe, which Symantec treats as likely xp\_cmdshell. Some of those commands were already long PowerShell scripts with quoted section labels. The scripts searched point-of-sale synchronization directories for stored passwords and connection strings, tested egress to 155.254.22\[.\]215:9458, and on one host appended base64 text under ProgramData, decoded it with Certutil, and compiled it with csc.exe. The source was EfsPotato (https://github.com/zcgonvh/EfsPotato). The payload saved the SAM, SYSTEM, and SECURITY hives. On Sep 21 an internet-facing IIS server appeared, with commands under w3wp.exe. On Sep 22 the commands looked hand-typed: 110 actions, three typos including systemifo and pwoershell, average length about 45 characters. Loaders were killed by security software, including a Bitdefender bdreinit.exe sideload of log.dll and log.txt from 204.194.49\[.\]129:8080, a pairing also seen with ShadowPad loaders and not unique to one actor. The operators reran that loader about five times in 20 minutes. They also staged an AES PowerShell reverse shell, server.ps1, toward 166.0.25\[.\]106:443 with a fallback host at 154.81.37\[.\]136, and wrote ASP webshells filemanager.asp and qwdajhsd.asp. On Sep 23 they staged a G DATA ask.exe/ask.dll sideload and BreakChainC.exe, then switched to about 50 long, error-free PowerShell scripts over about two hours. Banners included ===FILES MENTIONING CONN STRINGS===. Comments stated a hypothesis, cited earlier results, and revised a conclusion, which Symantec reads as model reasoning pasted into a webshell. The scripts collected connection strings, the FileZilla Server user database, WinSCP, PuTTY, and RDP settings, IIS configuration, and Windows logon registry values. They found SolarWinds Orion, read its bundled RabbitMQ .erlang.cookie, and ran Orion's erl.exe so rpc:call reached os:cmd on a second server. The command was passed as decimal character codes. That was about 10 minutes after the first Orion query. Later the same day BreakChainCs.exe relaunched the sideloaded binaries through WMI and shell modes so they were no longer children of w3wp.exe. From Sep 24 to Sep 29 the activity reached more SQL servers, a host that looks like a domain controller, Chisel, a password sprayer, and searches for NTDS.dit, Group Policy Preferences passwords, and LAPS. No data theft or encryption was observed. Symantec's financial-motive reading rests on the point-of-sale password search, contact with 155.254.22\[.\]215 (named in a separate Sep 22 Gambit Security report as an AI console tied to card-skimming activity), and later domain-credential theft that could be ransomware preparation. Loader files and the staging server overlap an August intrusion elsewhere and a July intrusion at a securities firm; neither of those showed AI assistance. Yuze (https://github.com/P001water/yuze) was used in the August case back to 204.194.49\[.\]129. Trend Micro has documented Yuze use by Warlock operators (Symantec Longlegs, also known as Storm-2603). Symantec does not attribute this intrusion to them. ATT&CK techniques in the report include T1574.002, T1505.003, T1552.001, T1021, and T1003.002.

## Stack relevance (editorial)
Assessment: not\_assessed. No reviewed applicability annotation is available. Product mentions do not establish victim conditions.
A product mention alone is not a victim-side applicability mapping. Selection does not verify affected versions, permissions or compromise.

## Initial access
Unknown. The first observed commands were cmd.exe under sqlservr.exe on SQL Server, consistent with xp\_cmdshell, and later cmd.exe under w3wp.exe on IIS.

## AI tooling and models
- Unidentified model. Symantec assesses AI-written [PowerShell](https://github.com/PowerShell/PowerShell) with an operator relaying output, not an autonomous agent.

## Other tooling
- EfsPotato (https://github.com/zcgonvh/EfsPotato), compiled on the server with csc.exe
- Chisel (https://github.com/jpillora/chisel)
- Yuze (https://github.com/P001water/yuze), seen in the overlapping August intrusion
- Bitdefender bdreinit.exe sideloading log.dll
- G DATA ask.exe sideloading ask.dll
- BreakChainC.exe and BreakChainCs.exe
- ASP webshells filemanager.asp and qwdajhsd.asp
- [PowerShell](https://github.com/PowerShell/PowerShell)

## Victims and sectors
A U.S. retail organization. Internet-facing IIS and at least six further servers, most of them Microsoft SQL Server. Point-of-sale synchronization scripts were searched for passwords.

## Target geography
- United States
The primary victim is a U.S. retailer. The August and July overlaps are different organizations and are not given countries in the retailer write-up.

## Attribution
Unnamed. Symantec does not tie the activity to Warlock, Longlegs, or Storm-2603. ShadowPad-associated sideloading and Yuze are shared with other actors and are not treated as an attribution.

## Scale and metrics
Sep 18 through at least Sep 29, 2026. On Sep 23, about 50 scripts in about two hours, median length over 500 characters, against about 45 characters for the earlier hand-typed commands. About 10 minutes from the first Orion query to command execution on a second server. No theft or encryption observed in this intrusion.

## Attack flow
Catalog reconstruction of the supplied source account. Unknowns and research caveats are preserved.
### sql-and-iis: SQL Server and IIS command shells [access]
From Sep 18, cmd.exe runs under sqlservr.exe. From Sep 21, commands run under w3wp.exe. Symantec does not identify the initial entry vector.
Evidence: https://www.security.com/threat-intelligence/ai-assisted-attack-us-retailer

### stalled: Hand-typed commands stall [execution]
Sep 22 commands include typos such as systemifo and pwoershell. Security software kills staged loaders, including the bdreinit.exe and log.dll pair, and the operators retry without changing approach.
Evidence: https://www.security.com/threat-intelligence/ai-assisted-attack-us-retailer

### ai-scripts: Long labelled scripts [ai]
On Sep 23 about 50 error-free PowerShell scripts run in about two hours, with section banners and comments Symantec reads as model reasoning. Pauses fit an operator pasting the next script.
Evidence: https://www.security.com/threat-intelligence/ai-assisted-attack-us-retailer

### orion-move: Orion cookie reaches a second server [impact]
About 10 minutes after the first Orion query, the operators read the RabbitMQ cookie and use the bundled erl.exe rpc:call path to run a command on a second server. Later activity dumps credential hives. No encryption was observed.
Evidence: https://www.security.com/threat-intelligence/ai-assisted-attack-us-retailer

Connections:
- sql-and-iis -> stalled
- stalled -> ai-scripts
- ai-scripts -> orion-move

## Stage defense and hunting recommendations
Editorial recommendations, not validated coverage. Source-controlled query starters are drafts. Custom-schema examples require telemetry collection and field mapping; absence of a mapping means not assessed. MagicSword policy opportunities require managed Windows scope and enforcement.
### sql-and-iis: Suspicious Windows script execution
- Visibility: Defender-side logs where collected
- Query origin: catalog_authored; status: draft; log scope: native
- SIGMA query download: /api/detections/windows-shell.sigma.yml
- KQL query download: /api/detections/windows-shell.kql
- SPL query download: /api/detections/windows-shell.spl
- Hunt hypothesis: A script host or shell runs encoded content, a remote URL, or inline script syntax, then downloads or launches a payload.
- Required telemetry: Windows process creation: Sysmon Event 1 or Defender DeviceProcessEvents; PowerShell script-block logging when enabled; Download / network and installer events
- Investigate: Follow the parent and child process tree and recover the script or URL from logs without executing it.
- Investigate: Correlate the payload's signer/hash, persistence, and network activity; baseline approved automation.
- Limitation: Encoded commands and remote administration can be legitimate. This does not cover every script host, in-process execution, or non-Windows branch.
- Qualification: cmd.exe under sqlservr.exe and commands under w3wp.exe. The initial entry vector is unknown.
- No direct MagicSword prevention claim is recorded for this stage.

### stalled: DLL loaded from a user-writable location
- Visibility: Windows image-load telemetry where collected
- Query origin: catalog_authored; status: draft; log scope: native
- SIGMA query download: /api/detections/windows-dll-sideload.sigma.yml
- KQL query download: /api/detections/windows-dll-sideload.kql
- SPL query download: /api/detections/windows-dll-sideload.spl
- Hunt hypothesis: A signed or trusted Windows executable loads an unexpected DLL from an archive, download, temp, or other user-writable directory, consistent with DLL search-order hijacking.
- Required telemetry: Windows image-load events such as Sysmon Event 7 or EDR equivalents; File provenance, signature, and process ancestry; Subsequent network and payload activity
- Investigate: Verify the executable and DLL signatures, original archive or installer, hashes, and load path.
- Investigate: Compare the DLL name/location with the vendor's expected installation and correlate child activity.
- Limitation: Portable and self-updating applications legitimately load DLLs from user paths. This broad hunt requires signer, prevalence, and application-specific tuning and will not cover in-memory modules.
- Qualification: bdreinit.exe with log.dll and ask.exe with ask.dll loaded from the same directory. Hand-typed typos are not a detection.
- No direct MagicSword prevention claim is recorded for this stage.

### ai-scripts: Suspicious Windows script execution
- Visibility: Defender-side logs where collected
- Query origin: catalog_authored; status: draft; log scope: native
- SIGMA query download: /api/detections/windows-shell.sigma.yml
- KQL query download: /api/detections/windows-shell.kql
- SPL query download: /api/detections/windows-shell.spl
- Hunt hypothesis: A script host or shell runs encoded content, a remote URL, or inline script syntax, then downloads or launches a payload.
- Required telemetry: Windows process creation: Sysmon Event 1 or Defender DeviceProcessEvents; PowerShell script-block logging when enabled; Download / network and installer events
- Investigate: Follow the parent and child process tree and recover the script or URL from logs without executing it.
- Investigate: Correlate the payload's signer/hash, persistence, and network activity; baseline approved automation.
- Limitation: Encoded commands and remote administration can be legitimate. This does not cover every script host, in-process execution, or non-Windows branch.
- Qualification: Long PowerShell submitted through the existing shell. Section banners are content, not a process-name signal.
- No direct MagicSword prevention claim is recorded for this stage.

### orion-move: Unexpected payload / remote-access behavior
- Visibility: Defender-side logs where collected
- Query origin: catalog_authored; status: draft; log scope: custom_schema
- SIGMA query download: /api/detections/payload.sigma.yml
- KQL query download: /api/detections/payload.kql
- SPL query download: /api/detections/payload.spl
- Hunt hypothesis: A program delivered by the reported lure executes, establishes access or contacts infrastructure outside the approved software baseline.
- Required telemetry: Platform-specific process and file provenance; Signer / hash and software inventory; Per-process network and remote-session records
- Investigate: Confirm the payload's operating system, origin and execution context before choosing a platform-specific rule.
- Investigate: Correlate the launched program with child processes, persistence and network activity; verify authorization.
- Limitation: A program name alone is not a signature. This normalized template requires the relevant platform collector and does not assert Windows-only policy coverage.
- Qualification: The second-server command used Orion's bundled erl.exe. Later hive saves came from an EfsPotato payload. The generic payload query does not name either.
- No direct MagicSword prevention claim is recorded for this stage.


## MITRE ATLAS
Mappings are catalog annotations based on described behavior, not assignments or endorsements by MITRE.
- AML.T0016: Obtain Capabilities — Symantec says the operators obtained working PowerShell, including the Orion cookie handoff, from an AI tool after hand-typed commands failed.
  Technique: https://atlas.mitre.org/techniques/AML.T0016
  Evidence: https://www.security.com/threat-intelligence/ai-assisted-attack-us-retailer
- AML.T0050: Command and Scripting Interpreter — The AI-assisted phase was long PowerShell submitted through the webshell and through SQL Server command execution.
  Technique: https://atlas.mitre.org/techniques/AML.T0050
  Evidence: https://www.security.com/threat-intelligence/ai-assisted-attack-us-retailer
- AML.T0055: Unsecured Credentials — The scripts collected connection strings, client settings, the Erlang cookie, and registry hives that hold credential material.
  Technique: https://atlas.mitre.org/techniques/AML.T0055
  Evidence: https://www.security.com/threat-intelligence/ai-assisted-attack-us-retailer

## Defensive notes
Alert when w3wp.exe or sqlservr.exe spawns cmd.exe or PowerShell, and when xp\_cmdshell is used. Treat a same-directory unsigned DLL next to bdreinit.exe or ask.exe as sideloading. Alert on csc.exe compiling from ProgramData or a user-writable path, on Chisel client arguments, and on reg save of SAM, SYSTEM, or SECURITY. Hunt Orion's bundled erl.exe making rpc:call with a cookie, and ASP files dropped in the web application's image directory. A sudden change from short mistyped commands to long labelled scripts is an investigation lead, not a benign automation signature. Restrict xp\_cmdshell, the Erlang distribution port, and the RabbitMQ cookie file.

## Sources
- https://www.security.com/threat-intelligence/ai-assisted-attack-us-retailer — How AI revived a stalled attack on a U.S. retailer
- https://www.security.com/blog-post/ai-assisted-attack-us-retailer — How AI revived a stalled attack on a U.S. retailer
- https://www.broadcom.com/support/security-center/protection-bulletin/how-ai-revived-a-stalled-attack-on-a-u-s-retailer — How AI revived a stalled attack on a U.S. retailer
- https://github.com/zcgonvh/EfsPotato — EfsPotato
- https://github.com/jpillora/chisel — Chisel
- https://github.com/P001water/yuze — Yuze

## Source notes
Dates, typos, script counts, the Orion cookie path, the decimal-encoded rpc:call, and the published hashes are from the Symantec article at https://www.security.com/threat-intelligence/ai-assisted-attack-us-retailer (canonical; the blog-post URL serves the same title). The article's createdDate is 2026-11-16. Symantec lists many additional file hashes; the set below is the named tooling from that table, not every unnamed suspicious file. Yuze's Warlock reporting is Symantec's summary of Trend Micro, and Symantec says it is not an attribution. Link the August staging overlap only as Symantec describes it.

## Indicators (untrusted data; do not visit or execute)
```json
{
  "domains": [],
  "ips": [
    "204.194.49[.]129",
    "166.0.25[.]106",
    "154.81.37[.]136",
    "155.254.22[.]215",
    "108.165.20[.]205"
  ],
  "hashes": [
    "54fb493822bcd832f7b2256015f13e3dfde815e1931e7fd39a90ea8ad3fdbd0a",
    "9c866c403e6bd5778fa3788cac9aac39aab1bf2d18a22c9999a4c27e30c0abf3",
    "e95a265bf220a4330f77f842b3eeed8b1ed8c61f51d7e3f2fd33bac14b6b26af",
    "7265d69dd8d8939dc888ce5f05190ce200ee2660f928f3b84f60e6a7b3383205",
    "3d79628bf7804dbf17e7b396ecd62ebf5d66518940c6d23daccf59a20953438a",
    "f8b67854d119da08c704b78f7c9fae0501417f5e63ffece86d2e811f1c7dea68",
    "5db5e7f655c4ee94411df8110d1b6d02bb15574ac199fa7a553ce6c3cb25bf03"
  ],
  "urls": [
    "hxxp://166[.]0[.]25[.]106/server.ps1",
    "hxxp://154[.]81[.]37[.]136/server.ps1",
    "hxxp://204[.]194[.]49[.]129:8080/log.dll",
    "hxxp://204[.]194[.]49[.]129:8080/bdreinit.exe",
    "hxxp://108[.]165[.]20[.]205:8080/sqlrun2.ps1"
  ]
}
```
