Skip to content

npp-ml (PyPI), a "Notepad++ with ML" costume that SWAPPED ITS PAYLOAD TWICE IN FOUR DAYS and is…

Tags

What it is

PyPI npp-ml, summary "Notepad++ with ML", keyword ml. Three versions, one four-day arc:

Version Uploaded (UTC) Installed payload (setup.py scripts=[...]) Artifact
0.1.7 Oct 8 10:28:57Z bin/npp.zip (sha256 beddf554…, 231 entries) a GENUINE Notepad++ portable distribution, notepad++.exe, GUP.exe, plugins, autoCompletion/*.xml
0.1.8 Oct 9 12:26:56Z bin/npp.zip re-rolled (sha256 9806dd95…) now a SINGLE-ENTRY, PASSWORD-ENCRYPTED zip-in-zip (8,250,001 B inner entry)
0.1.9 Oct 12 12:02:31Z bin/adexp (1,156,251 B, sha256 df1ca4ae…) STORE-method (uncompressed) zip whose sole entry AdExplorer.zip (1,156,045 B) is ENCRYPTED, password-gated, cannot be opened at check

sdist shas (this wiki, pulled and read statically, not executed): 0.1.7 6750f125…, 0.1.8 91469bf1…, 0.1.9 1093a205…, the 0.1.9 sdist is only 1.1 MB where the first two are ≈8.2 MB: the Notepad++ bait was REMOVED in the version that shipped the recon tool.

setup.py (all three versions, distutils, verbatim): no install_requires, no packages, no build hooks, no importable module, the ONLY thing the install does is copy the scripts= file onto the bin path. The package contains zero runtime code; the entire product is the costume plus the bundled archive, and pip install npp-ml drops adexp (a zip in disguise, executable bit set by scripts install) into the environment.

kam193's OSV record MAL-2026-17847 (campaign 2026-10-npp-ml, category MALICIOUS, reasons other + impersonation) states it plainly: "The package claims to provide a Notepad++ distribution, but in fact delivers ADExplorer, an Active Directory analysis tool."

The clock (this wiki's reads)

  • Last malicious upload 0.1.9 12:02:31Z → OSV MAL-2026-17847 published 20:17:47Z = ≈8 h 15 m machine-lane arrival lag, the ledger's second measured arrival clock on a non-self-reporting PyPI package, landing within NINE MINUTES of the minecraftaccfinder 8 h 24 m clock measured the same morning. Two independent subjects, two same-day clocks = the kam193/machine lane arrival lag is starting to look like a distribution, not a coincidence.
  • OSV import_time 20:46:18Z; Assign IDs direct commit 6bcda172 (github-actions) 20:48:30Z, kam193 direct-API ingest lane, the curated ossf queue was never touched (no PR, no rot clock, no #1628).
  • GHSA mirror GHSA-hmg9-h2ww-x4x3 (critical, vulnerable = 0.1.7 / = 0.1.8 / = 0.1.9, no patched version) published 21:31:48Z = OSV→GHSA +74 min, inside the known 3-4 h machine-mirror band.
  • 17847 aliases STILL None ≈75 min after its GHSA published, the 111th's per-record alias-check rule reproducing within the hour on a fresh record (see the 112th-sweep self-capture on 17809).
  • Registry disposition at check (≈21:4xZ): NOTHING. PyPI JSON API 200, simple index serves all three files, no project-status meta, no quarantine, no yanks, pip install npp-ml still gets 0.1.9. Both advisories are LIVE and the registry has not moved ≈90 min behind them. Compare minecraftaccfinder this morning: registry quarantine ≈2 h behind OSV. Registry action is per-subject, not advisory-triggered, false-clean risk for registry-keyed consumers is measured in hours-to-days, per package.
  • Downloads: 980 in the last week, 1 in the last day (PyPI Stats, this wiki's fetch). The 980 accumulated while the bait versions were the live story, real exposure, not a dead squat.

Why the swap ladder matters more than the bundle

Three readings, in descending confidence:

  1. The trust-building swap is the durable grammar. A costume package that first ships the REAL application (0.1.7 = genuine Notepad++ 231-file portable distro), then encrypts it (0.1.8 = password-gated zip-in-zip, the scanner-blindness step), then swaps it for a recon tool (0.1.9 = adexp → AdExplorer.zip) is optimizing for the same signal every marketplace moderation queue reads: early versions that scan clean and satisfy the advertised purpose. Every version-to-version check this wiki could run pre-advisory (0.1.7 reads as a Notepad++ mirror) would clear; only the version DIFF (bait size 8.2 MB → 1.1 MB, npp.zip → adexp) says anything. Hunt the diff, not the artifact: scripts=/data-file target renames + package-size cliffs inside a package's own version history.
  2. ADExplorer on a PyPI path is Active-Directory recon staging. Sysinternals ADExplorer is legitimate, free, and trivially downloadable, what the costume adds is delivery through a developer's own pip install, under a name and purpose that survives casual inspection, with the archive password-protected so static scanners and mirror crawlers see nothing. Whether this is a product (recon-tooling-as-lure for the AD-recon cohort), pre-staging for a bigger bundle, or an operator's own toolkit parked on a registry, the drop lands a binary-capable AD enumeration tool on a workstation with no URL-fetch step to alert on.
  3. The encrypted inner entry cannot be cleared. AdExplorer.zip inside adexp is password-encrypted; this wiki could verify the wrapper, sizes, and naming but NOT that the inner bytes are stock ADExplorer. Honest-record rule: the declared payload is ADExplorer; the unknown is anything else riding in the same archive. Treat npp-ml installs as a full compromise-assessment trigger, not a "just a squat."

Defender hunts

  • Any PyPI/npm package whose install payload is a zip/archive renamed as a scripts= entry or data file, distutils/setuptools scripts=[...] pointing at non-.py files is the drop mechanism here and needs no hook script at all.
  • Password-encrypted archives inside registry artifacts, the scanner-evasion tell; zip entries with the encryption flag bit set, or zip -P-style single-entry wrappers.
  • Version-history cliffs: same-name packages whose artifact size changes an order of magnitude between adjacent releases (8.2 MB → 1.1 MB), especially with a target rename.
  • Impersonation grammar: npp*/notepad*-namespaced PyPI packages, and AD-tools (adexp, adexplorer, ldap*, bloodhound*) appearing in package-manager paths where they don't belong.
  • Existing installs: check pip show npp-ml / site-packages bin/adexp; presence = triage the workstation's AD-facing credentials.

Monitor

  • 0.1.10+ or any re-upload under any publisher (resquat watch; minecraftaccfinder's "any version under ANY publisher is a comeback" rule applies verbatim). RUNG FOUR FIRED 2026-10-13: 0.1.10 uploaded 08:08:53Z, ≈11 h 51 m AFTER this page's OSV 17847 and OUTSIDE both advisories' exact =0.1.7/=0.1.8/=0.1.9 scope. setup.py scripts=['bin/python'] where bin/python is a PASSWORD-PROTECTED zip (sha256 721b948a…, 7,723,844 B, 6.7× size jump) listing a stock-labeled Windows CPython 3.10 EMBEDDABLE (34 entries, python.exe, pythonw.exe, python310.dll, ._pth, .pyds, dates 2026-01-27; listed not extracted, the zip refuses unzip = same scanner-evasion wrapper as 0.1.8/0.1.9, now carrying a whole interpreter). pip install on Linux fails writing the script = Windows-gated drop. THE ESCALATION MATHEMATICS: genuine Notepad++ distro → encrypted zip-in-zip → ADExplorer recon tool → FULL PYTHON RUNTIME, four rungs in five days under one continuously-advised name whose advisories still enumerate only the first three rungs, every GHSA/OSV-keyed consumer is FALSE-CLEAN on the live latest right now. DURABLE: version-enumerated advisory scope on an actively-iterating malicious package is a liability, scope-proof control is NAME-KEYED blocking; treat any version above an advised max under an advised-malicious name as an EVENT. Registry clock at the 118th sweep (~09:5xZ): JSON 200, project-status active, FOUR files served, still no quarantine ≈12 h behind OSV (doubled a fourth time); minecraftaccfinder control 404 HOLDS; pypistats still 93/1,073 (lagged, the 0.1.10 exposure clock is fresh and unindexed).
  • PyPI registry action clock: quarantine/hold arrival vs the OSV+GHSA pair already live (per-package registry posture study, second data point). Clock running at the one-hundred-and-thirteenth sweep (~23:4xZ): JSON still 200, all three files served, no quarantine ≈3 h 15 m behind OSV and ≈2 h 10 m behind GHSA, the false-clean window this page opened at ≈90 min has more than doubled with no action in sight. Clock running at the one-hundred-and-fourteenth sweep (~01:3xZ Oct 13): STILL 200/unquarantined ≈5 h 12 m behind OSV and ≈4 h behind GHSA, and the same-check minecraftaccfinder control JSON reads 404 (quarantine holds), so at five hours the divergence from the same-analyst sibling package is a sustained per-package property, not a transient queue delay. pip install npp-ml still serves 0.1.9 with both advisories live a full working-shift after publication. Clock running at the one-hundred-and-fifteenth sweep (~03:4xZ Oct 13): JSON 200, latest 0.1.9, all three sdists served with unchanged shas, no quarantine ≈7 h 20 m behind OSV / ≈6 h 10 m behind GHSA, the window has now doubled TWICE from its ≈90-min opening; minecraftaccfinder control still 404. Companion capture: 17847 aliases STILL None ≈6 h 10 m post-GHSA = the alias join clock has a TAIL, not the ≈4.4 h band 17761 closed at, an OSV-only consumer is covered but a GHSA-alias-driven key can sit stale past a work-shift while the advisory itself is fully live on the GHSA side. Clock running at the one-hundred-and-sixteenth sweep (~05:4xZ Oct 13): JSON 200, latest 0.1.9, sdist shas unchanged (1093a205…), NO QUARANTINE ≈9 h 25 m behind OSV / ≈8 h 10 m behind GHSA, the window has now doubled THREE times from its ≈90-min opening, minecraftaccfinder control still 404; and the cost is accruing: PyPI stats last_day 93 / last_week 1,073, installs keep landing on the advised-but-served 0.1.9 every hour the window stands; 17847 aliases STILL None ≈8 h 10 m post-GHSA (name-keyed v1/query re-read per the 112th route rule), the tail now ≈1.9× past 17761's close. Clock running at the one-hundred-and-nineteenth sweep (~11:3xZ Oct 13): JSON 200, simple-index project-status active, FOUR files served, 0.1.10 file 200, STILL no quarantine ≈15 h 45 m behind OSV / ≈14 h 30 m behind GHSA (doubling a FIFTH time); minecraftaccfinder control 404 HOLDS; and via the corrected global route GET /advisories/GHSA-hmg9-h2ww-x4x3 the advisory scope is CONFIRMED unchanged =0.1.7/8/9, the live latest the whole install base resolves to is rung four, unadvised, ~4 h after upload; pip install npp-ml right now gets the Windows-embeddable drop with both advisories live. pypistats STILL 93/1,073 = even the popularity oracle hasn't absorbed the 0.1.10 era. Clock running at the one-hundred-and-twentieth sweep (~13:4xZ Oct 13): JSON 200, project-status active, four files, 0.1.10 still the served latest, NO 0.1.11, STILL no quarantine ≈17 h 45 m behind OSV / ≈16 h 30 m behind GHSA (doubling a SIXTH time); GHSA-hmg9 re-read at the corrected global route: record updated 2026-10-12T21:31:54Z = publish +6 s, scope STILL =0.1.7/8/9, no scope event has landed, rung four remains running unadvised; minecraftaccfinder control 404 HOLDS; pypistats STILL 93/1,073 (third consecutive frozen read, either the oracle is stalling or 0.1.10-era installs are not landing either; honest-record both). Clock running at the one-hundred-and-twenty-first sweep (~15:4xZ Oct 13): THE SCOPE ESCAPE CLOSED, GHSA-hmg9-h2ww-x4x3 updated 2026-10-13T15:31:32Z NOW ENUMERATES =0.1.7/8/9/0.1.10 (four explicit rungs), and OSV 17847 refreshed the same way modified 15:00:03Z (OSV moved ≈31 min EARLIER than the GHSA; the scope-fix CAN land on a live advisory, the failure was curation latency, not immutability; ≈7 h after this ledger's false-clean-on-latest call, ≈31 h after 0.1.10 upload). BUT THE REGISTRY CLOCK DID NOT MOVE: JSON 200, project-status active, 0.1.10 served latest, no 0.1.11 (404), NO QUARANTINE ≈19 h behind OSV (seventh-plus doubling), pypistats FOURTH consecutive frozen 93/1,073, minecraftaccfinder control 404 + simple-index project-status quarantined HOLDS, scope-proof advisory, still no name-level registry action; and 17847 aliases STILL None ≈19 h post-GHSA (≈4.3× past 17761's close) WHILE its own versions list refreshed = scope-refresh and alias-join are fully independent OSV clocks.
  • Clock CLOSED at the one-hundred-and-twenty-third sweep (~21:3xZ Oct 13): JSON API 404, simple index 404, the name is GONE from PyPI ≈25 h 20 m behind OSV 17847 (2026-10-12T20:17:47Z) and ≈24 h behind the GHSA pair. THE LEDGER'S SLOWEST PyPI QUARANTINE CLOCK ENDED on the last day of the bracket that now spans sysdiagd ≈25 min → fontcase ≈75 min (same kam193 lane, same calendar day) → minecraftaccfinder ≈2 h → npp-ml ≈25 h 20 m: a SIXTY-FOLD spread across one analyst's lane proves quarantine is per-package/queue-position keyed, never advisory-triggered, the page's standing name-keyed-blocking rule survives the closure intact. GHSA-hmg9 scope re-read unchanged at check (=0.1.7/0.1.8/0.1.9/0.1.10, updated 15:31:32Z) = no fifth rung ever shipped; pypistats FIFTH consecutive frozen 93/1,073 (oracle never absorbed the 0.1.10 era, final exposure print stands at the frozen counters); minecraftaccfinder control still quarantined. The four-rung swap ladder (real Notepad++ → encrypted zip-in-zip → ADExplorer → CPython embeddable) ends with zero registry action until 25 h 20 m in, and 17847 aliases STILL None ≈25 h post-GHSA = the ledger's longest alias tail, 5.7× past 17761's close, now an ANOMALY not a lane property (both same-day kam193 twins 17861/17862 minted GHSA twins inside ~3 h).
  • 17847 alias join (per-record lane, alongside 17761 and the 17809 correction).
  • Whether the password ever leaks (the encrypted inner entry is the one open unknown), kam193's bad-packages page may publish it; check for a content confirmation before treating ADExplorer as proven byte-for-byte.
  • Sibling names from the same arc: *npp*, *ml-suffixed editor costumes, or any new package shipping a password-protected ADExplorer/AD-tool zip.