Lumma Stealer's PNG stego loader

Summary

Huntress traced incidents in which a counterfeit human-verification page, served from Cloudflare Pages subdomains, placed an mshta command on the visitor's clipboard and told them to run it from the Windows Run box. The page held its second-stage script XOR-encrypted in a variable and injected it through a blob URL, which frustrates string matching. mshta retrieved a JScript file from a bare IP address written with a hex-encoded second octet; that ran PowerShell in memory, which decrypted a .NET assembly and loaded it reflectively. The assembly carried an AES-encrypted PNG in its manifest resources and rebuilt Donut-packed shellcode out of the red channel of the pixels, XORing each value against 114. Injection into explorer.exe followed, and the unpacked payload was Lumma Stealer. Huntress released a refreshed configuration extractor for the samples it recovered.

Overlap with other campaigns

We compared the 19 indicators in this record — covering 11 distinct registrable domains — against the 43 other published campaigns in this corpus.

• AMOS via a fake Spectrum page

  2026-08-07

  No indicator is shared, but the two records use the same registrable domains.

  Shared registrable domain (1)

  • pages[.]devdomain, shown defanged for safety: pages dot dev

A shared indicator is the same record cited by both campaigns. A shared registrable domain is weaker evidence and is counted separately, because two campaigns can share a registrable simply by renting subdomains from the same platform.

What changed since publication

The accounts above describe these indicators as they were on the day each was written. We re-check them ourselves and record every check, including the ones that find nothing — which is the only reason the gaps below can be stated rather than hidden.

Indicators checked

8 of 19

Checks recorded

16

Checks since publication

16

Never checked

11

Observation window 2026-08-10 to 2026-08-22. Checks are sampled, not continuous; the sampling policy is on /methodology. 11 indicators have never been checked by us at all.

Nothing has changed state between our first check and our most recent one. A domain that still resolves is not necessarily still serving anything — registrar hold pages answer DNS indefinitely, which is why the state above distinguishes “parked or suspended” from “resolving”.

Execution mechanisms

How the pasted command actually ran, as the published accounts describe it. Every row under “Mechanism” is one this campaign records; “Across the corpus” is the share of the 44 published campaigns in this corpus that record the same one — a baseline we can only state because we hold the whole set.

All 5 of these mechanisms appear together in 2 of the 44 campaigns on record, this one included. We record a mechanism only where a source describes one: an absent mechanism means nobody wrote it down, not that it did not happen.

Indicators

19 indicators, each present verbatim in at least one of the sources above. Values are shown defanged and are not links to the hosts — the link goes to our page for that indicator.

“Assessment” is who controls the host, not how dangerous it is. Many of these are compromised legitimate sites whose owners are victims, and they are excluded from every blocklist file we publish.