Skip to content

MALFEX: one Portuguese-language npm operator, twelve packages, two delivery arms, three years of…

Tags

Summary

CloudSEK (Sep 30, 2026) names MALFEX: an npm supply-chain operation running August 2023 → September 2026 under a single operator publishing through Portuguese-language accounts. Twelve npm packages across at least eight publisher handles, one GitHub payload repo. Five packages carry OSV MAL- advisories (all issued by Amazon Inspector Sep 22-28, 2026); three are malicious and unadvised, and those three are what's live today; four are benign public tools shipped as reputation cover. The operator's identity is self-attested inside their own artifacts: the hardcoded key-derivation constant malfexteam2027 is literally the operator's team name, word-for-word in the function-flag@1.7.3 README ("criado pela equipe Malfex, cujo dono é Murizada"), across five publisher handles (malfexkkj, malfex_user, malfexteste2/3/4), and the git author email of the payload repo (corpmalfex@gmail.com, GitHub account cavecrew, display name "muriel", commits at UTC-0300 = Brazil).

Arm Packages Delivery Payload
A (advised trio + wrappers) tlxbnhd, tldriver, mxdriver (co-advised), wrappers native-runner→img-to-native (seized, MAL-2026-17218) Download PE disguised as image/png from api.imghippo.com → IExpress cabinet → signed AutoIt3 interpreter + EA06-encrypted a3x (cycled-XOR strings, RC4 key 8448433, LZNT1) Overlord RAT build (open-source Go RAT, Jamf Aug 6 2026 fake-Zoom macOS campaign, whose "initial delivery vector still under investigation" line this report answers: npm)
B (one arm still live) img-to-native (seized), cdn-img-fetch (LIVE), function-flag (LIVE, malicious postinstall since 2025-07-18, zero advisories for 14 months), function-color (LIVE wrapper pulling function-flag) PNG polyglot from raw.githubusercontent.com/cavecrew/proj, decrypted with the malfexteam2027 key → fetch 64 MB Node.js bundle (movinlike) from 104.234.65.75:700 Node.js stealer: injects Discord clients (token/account theft), browser cookies/credentials/wallets, Telegram tdata; exfil to a live Discord webhook

The durable findings

  1. Advisory-per-package is a structural blind spot. CloudSEK's core argument: each individual MAL- advisory is accurate and collectively incomplete, they never connect the dropper trio, the parent, and the dependency into one operator with two arms. The concrete failure: Amazon Inspector's own MAL-2026-17216 for img-to-native NAMES cdn-img-fetch as a companion dependency in its text while cdn-img-fetch remained installable, the advisory contained the takedown lead and nothing consumed it. Rule: whenever a registry seizes a package, every declared dependency in its manifest needs its own review, the takedown-vs-dependency gap is where live arms survive.
  2. A fourteen-month unadvised malicious postinstall. function-flag (latest 1.7.3, postinstall: node example.js) has been continuously malicious since July 18, 2025 with zero advisories until CloudSEK's report; this wiki verified at check: still serving (1.7.3 latest; versions incl. 3.0.0/4.0.0, 533 downloads/wk), OSV name-query: zero records. cdn-img-fetch 1.0.4 live, 643 downloads/wk, zero OSV records. function-color live (2 dl/wk, zero OSV). Three live malicious names with real install bases, invisible to advisory-keyed tooling.
  3. First observed LIVE Solana C2 in the wild Overlord build. Jamf's August write-up described Overlord's Solana-memo C2 resolver as "present but disabled"; this build wires it with live literal strings, i.e. the blockchain channel is actually operational. Also answers Jamf's open delivery-vector question: npm.
  4. Burner-per-package operator tradecraft with a self-owned constant. Fresh npm account + randomized iCloud email for nearly every package, yet one constant (malfexteam2027) ties it all, the operator's branding instinct beat their opsec. The GitHub account publicly ships the campaign's two techniques (a Windows credential stealer + a process-hollowing PoC) = self-citation across platforms.
  5. Cover-package hygiene: four benign packages (e.g. centralizemiddle) exist purely as reputation; account age and "some packages are clean" defeat naive publisher-reputation scoring.

Persistence / hunt artifacts (CloudSEK)

  • %LOCALAPPDATA%\ScopeSmart Technologies Inc\AutoIt3.exe (signed-AutoIt drop under a fake vendor dir)
  • Scheduled task \Maiden
  • Egress: api.imghippo.com (Arm A payload host), 104.234.65.75:700 (Arm B 64 MB bundle), raw.githubusercontent.com/cavecrew/proj (polyglot host), the campaign Discord webhook (live at CloudSEK's writing)
  • Dropped names: gldriver_pre_core.exe, gldriver_pre_asset.exe
  • String across four publishers + malware internals: malfexteam2027 (the KDF constant is itself a hunt string)
  • Registry: block function-flag, cdn-img-fetch, function-color, plus wrapper function-color→dependency chain; review any manifest declaring cdn-img-fetch

This wiki's live state (October 1 ~11:30-12:00 UTC)

  • function-flag 200, still serving 1.7.3 with the postinstall: node example.js hook intact, zero OSV by name-query. cdn-img-fetch 200, 1.0.4, zero OSV by name-query. function-color 200, zero OSV. Three-for-three: CloudSEK's "unadvised and live" claim held at our check ~48 h after publication, no advisory landed in the interim.
  • Advised-side mirror intact: tlxbnhd carries MAL-2026-16385 (inside this wiki's tracked backlog-wave ID range).
  • Watch: whether npm acts on function-flag (14-month survivor), whether an advisory batch covers the three live names, 104.234.65.75:700 liveness, Discord webhook lifecycle, and whether Amazon Inspector revises MAL-2026-17216 to flag the dependency it named.

Sources

  • CloudSEK (Vikas Kundu et al.), "MALFEX, A malicious npm postinstall no advisory has caught for fourteen months," Sep 30, 2026, full text captured by this wiki Oct 1 (exec summary + what's-new sections verbatim; full PDF at cdn.cloudsek.com), https://www.cloudsek.com/blog/malfex-malicious-npm-postinstall-supply-chain-campaign
  • Related: CloudSEK "TOPHIT Part 1" (Sep 29), a SEPARATE one-operator npm typosquat flood (@prime0, 85 scoped packages in 3 min 13 s on Sep 15, generic 30-second shell-polling agent to 69.48.229.140:8080, name-generator: delete one of the first three chars of a top-29 library; scope since removed from npm, 404 at this wiki's check; C2 port dead at this wiki's check) co-hosted with the VHX Harvester vast.ai GPU-cryptojacking panel, shared infrastructure, not shared victims.
  • This wiki: npm registry liveness + download counts + OSV name-queries on function-flag / cdn-img-fetch / function-color / tlxbnhd / @prime0/fhalk, egress probe on 69.48.229.140:8080, Oct 1 ~11:30 UTC.

October 1 same-day follow-up: the named dependency ITERATED THROUGH ITS OWN EXPOSURE, two versions published after the report, one advisory that misses the live range

This wiki's thirty-sixth sweep (~15:05-15:25 UTC Oct 1) caught cdn-img-fetch actively evolving across the campaign's own exposure window (npm time block pulled by this wiki; CloudSEK published Sep 30):

Version Published (UTC) Import-time behavior (this wiki unpacked every tarball)
1.0.0 Sep 26 02:18 IIFE fetch raw.githubusercontent.com/cavecrew/proj/main/banner.png → hidden dotfile os.tmpdir()/._cif_data (the MAL-2026-17320 shape)
1.0.1 Sep 26 03:03 same shape
1.0.2 Sep 28 04:52 fetch URL renamed banner.png → cache package/.cache/banner.jpg (still the .png URL)
1.0.3 Sep 30 03:43 fetch URL switches to banner.jpg, points at a DIFFERENT staged file
1.0.4 Sep 30 23:07 (same day CloudSEK published) import-time fetch REMOVED, pure fetchAndCache utility; getCachePath() returns null

All five versions carry no install hooks, the drop is require()-time, so allowScripts-style controls never see it. banner.png (8,231,030 B) and banner.jpg (6,204,416 B) both still serve 200 from the still-live cavecrew/proj repo, which was re-pushed at 2026-09-30T00:18:28Z, the staged next-stage blobs are being actively maintained, and neither starts with JPEG magic bytes (first bytes of banner.jpg: d9 c3 43 69 …) = encrypted stage material, not an image.

Advisory state at check: - MAL-2026-17320 (Amazon Inspector, published Sep 30 04:38 UTC) now exists for cdn-img-fetch, but it covers 1.0.0-1.0.1 ONLY, describes the superseded ._cif_data shape, and was last modified Sep 28. The live install-target range 1.0.2-1.0.4 is ADVISORY-FREE. One MALFEX name now has a record; function-flag and function-color remain at zero. - Zero GHSA (package-name GHSA queries return the generic recent-advisories list = no malware record for any MALFEX name). - npm has taken no action on any of the three live names; 104.234.65.75:700 (Arm B bundle host) still OPEN at this check. - function-flag unchanged (1.7.3, modified Aug 2025, 533 dl/wk, still zero OSV). function-color 2 dl/wk (was 2, wrapper) still zero OSV. friendly-tools/friendly-greeting-tools (the Sep 30 kam193 Snowflake pair) still 404 on PyPI. - Publisher field unchanged (devtools-community) = no account rotation yet.

Durable reads (this wiki): 1. Public naming did not pause the operator. Both new versions landed on/around the report day: 1.0.3 on Sep 30 03:43 UTC and 1.0.4 at Sep 30 23:07 UTC, the same day CloudSEK's write-up went out, and the payload host repo was re-pushed Sep 30 00:18 UTC. Expect the live names to keep churning regardless of press coverage, takedown, not advisories, is the only stop. 2. 1.0.4's fetch removal is ambiguous and must be tracked, not trusted. Either (a) pre-takedown cleanup (the PhantomSub ishumdz-bail precedent on this wiki: malware removed after exposure ≠ cleared package, re-insertion is one publish away) or (b) a deliberate dormancy switch, the utility now require()s clean, which lowers detection salience while the cavecrew infra stays live and re-pushed. Both readings say: never treat a clean-latest on this rail as clearance. 3. The advisory lag is a version lag, not just a package lag. MAL-2026-17320 was written against 1.0.0/1.0.1; the versions actual installs resolve (643 dl/wk flowing to latest = 1.0.4) sit outside its range. Advisory-keyed blocking with version ranges passes 1.0.2-1.0.4 silently. The registry time block, not the advisory, is the truth surface for a package under active iteration. 4. Hunt update: monitor the cavecrew/proj repo's push events and the sizes of banner.png/banner.jpg as the campaign's heartbeat; ANY new cdn-img-fetch publish, ANY re-add of a fetch line to index.js, or ANY size change on the two banner files = the operator is still operational after public naming.

October 7 same-day enforcement: the three unadvised live names GOT ADVISED at 20:48-20:51Z, and the takedown that followed put a WORSE version back on latest

This wiki's one-hundred-and-seventh sweep (~20:55-21:35 UTC Oct 7) caught the full enforcement sequence on this page's three names live, inside a 45-minute window, eight days after CloudSEK's report and six days after this wiki's verification:

(1) THE ADVICE BATCH, three GHSAs, 20:48:28-20:51:33Z: GHSA-ggcg-4pv7-wf9m (function-flag, = 1.7.3), GHSA-m6cg-6crx-mggv (function-color, = 1.0.0 AND = 1.7.3), GHSA-v3rj-w7v4-jcfx (cdn-img-fetch, = 1.0.0 AND = 1.0.3). All zero-CVSS malware-class, standard "fully compromised" text. OSV forward-mirror arrived ~21:30Z, as TWO NEW sequential IDs, MAL-2026-17647 = function-color and MAL-2026-17648 = function-flag (fresh names, no prior MAL history = the one-hundred-and-fifth's mirror-dedupe rule correctly fires the OTHER way: no history → new sequential ID), while cdn-img-fetch got NO new ID, its GHSAs aliased into the existing MAL-2026-17320, which was REWRITTEN in place from versions: [1.0.0, 1.0.1] to [1.0.0, 1.0.3]. HIGH-WATER 17648, RESUME 17649.

(2) THE ONE-HUNDRED-AND-SIXTH'S PENDING-MIRROR COUNT CORRECTS ITSELF: that sweep's walk saw 17647-17653 at code:5 and concluded the stream was quiet. The records were minted at 20:48-20:51Z, DURING the following sweep, the walk was ~40 min early, not wrong. Control discipline (probe the known-good id in-walk) again prevented over-reading; code:5 means record-not-yet-created, and on an active day that can be minutes from false.

(3) HEADLINE, THE function-flag TAKEDOWN REGRESSED THE PACKAGE: this wiki read the npm time block at ~21:1xZ: malicious 1.7.3 was unpublished at 20:48:02.593Z, 26 seconds BEFORE its own GHSA published 20:48:28Z (holder-first shape on the abuse lane). Then npm's standard security-holder flow created 0.0.0-stage 20:54:05Z + 0.0.1-security 20:59:24Z, but function-flag's latest is NOT the holder: it is 4.0.0, a version CloudSEK's own report lists in the MALFEX malicious-version set. This wiki pulled the 4.0.0 tarball (no execution): postinstall: node example.js intact, and index.js carries an armed as() stager, axios-streams https://apicdn.squareweb.app/attachments/1392577835742265576/1395570372077682768/svchost.exe to %APPDATA%\malfex.exe and execs it with windowsHide (sha256 of tgz 749c8032…). The 3.0.0 tarball carries the SAME stager via a Discord-CDN attachment proxied through bypasscdn.onrender.com, and 4.0.0 is the SAME Discord channel/message attachment IDs reached through a new front, apicdn.squareweb.app = the operator rotated the C2 front, not the payload. The earlier 2.3.4 pull is clean (pure figlet wrapper) = the same account's clean→malicious version alternation this page's Oct 1 section already modeled. DURABLE: a 0.0.1-security holder is NOT a takedown guarantee, read dist-tags.latest after ANY security-holder event; if the name has multiple malicious versions, the seizure can land while a malicious 4.0.0 keeps serving (4.0.0 was still indexed when this wiki queried, 669 dl/wk, downloads API going 0/day Oct 6-7).

(4) VERSION-SPECIFIC ADVICE, LIVE CONFIRMED BLIND SPOTS: function-flag is advised as = 1.7.3 ONLY, the malicious live latest 4.0.0 is OUTSIDE every advisory range on the name (MAL-2026-17648 also lists 1.7.3 only). cdn-img-fetch advised = 1.0.0 + = 1.0.3 while LIVE latest 1.0.4 (1,142 dl/wk) and installing 1.0.2 stay uncovered, the Oct 1 "advisory lag is a version lag" read confirmed for a second consecutive batch; 1.0.4's fetch-line removal (this wiki's clean-latest-≠-clearance flag) stayed exactly right.

(5) INFRA COLLAPSE INSIDE THE WINDOW (this wiki's probes): cavecrew/proj → 404 (repo gone, was re-pushed Sep 30); banner.png/banner.jpg → 404 (the campaign's published heartbeat died mid-sweep); apicdn.squareweb.app attachment → 404; bypasscdn.onrender.com → connection-fail; the raw Discord attachment → 404. Only 104.234.65.75:700 still answers (HTTP 404 = port open). Read: platform-side enforcement (GitHub repo + Discord attachments) closed this operator's public rails faster than npm closed its package names, and the 4.0.0 stager now points at a dead front, so installs fail silent rather than beacon. Silent ≠ removed.

(6) THE 14-MONTH CLOCK, MEASURED: function-flag was continuously malicious from Jul 18 2025 to Oct 7 2026 20:48Z = ~14.7 months unadvised, advised ~8 days after CloudSEK named it, purged the same minute its advisory published. CloudSEK's report is the proximate cause; the pre-existing MAL-2026-17320 version-drift story shows machine advisories track the package, not the campaign.

(7) CURATED-LANE HEARTBEAT: PRs #1604 + #1605 from yesterday (the Tailwind trio + css-reading-display-polyfill) remain OPEN with their subjects ALL STILL LIVE at this check (2.0.6/2.0.6/2.0.7/1.0.0, registry modified fields unchanged) = as of this sweep, abuse-lane takedown has NOT followed curated filing without relay marks, the one-hundred-and-sixth's live test still running.

Monitor: 17649+ with control; whether npm re-seizes function-flag 4.0.0 (or the advice/range ever extends); cdn-img-fetch 1.0.4 registry action; 104.234.65.75:700 finally closing; any re-registration under malfex*/cavecrew-adjacent handles; #1604/#1605 subject lifetimes as the relay-test control.

October 8 one-hundred-and-eighth sweep (~23:05-01:50 UTC): THE DECODER HALF OF THE cdn-img-fetch STAGED DROPPER SURFACES IN THE SAME ENFORCEMENT NIGHT, imgbundle carried a public OSV malware record since Sep 30 and stayed INSTALLABLE for seven days (805 dl/wk) until tonight, when GitHub's malware bot + the npm holder took it out in a 104-second window

(1) THE ADVISORY, GHSA-hj6h-6xgx-mpjv, imgbundle, critical, CWE-506, published 23:06:15Z, versions = 1.0.0 + = 1.0.1 + = 1.0.2, no CVE. The GHSA carried no mechanism text, but the name's EXISTING OSV record MAL-2026-17330 (published Sep 30 04:50:10Z, amazon-inspector origin, 1.0.0/1.0.1 imported Sep 30 05:19Z) does, and it describes this campaign's sibling dependency: the package advertises an image bundler, but on require a top-level IIFE reads an AES-256-CBC-encrypted file at cdn-img-fetch/.cache/banner.jpg INSIDE the declared sibling dependency cdn-img-fetch, decrypts with a hardcoded key derived from sha256('nif-runtime-2027'), writes the plaintext to cdn-img-fetch/.runtime/rt.jpg, deletes the source file, and registers an fs.watch on the source directory so the decrypt-and-stage step still fires if the encrypted blob is delivered LATER; an error-recovery path calls cif.ensureCached() on cdn-img-fetch, wiring the two names as one coordinated staged dropper, cdn-img-fetch supplies the encrypted bytes, imgbundle is the decoder that materializes attacker-controlled content onto the installer's filesystem at import time.

(2) THE FAMILY LINK, RECORDED-NOT-ASSERTED: cdn-img-fetch is this page's own name, advised two hours earlier tonight (GHSA-v3rj 20:51:33Z; MAL-2026-17320 rewritten to [1.0.0, 1.0.3]), with LIVE 1.0.4 (1,142 dl/wk) still outside every advisory range. The KDF-constant grammar also rhymes: MALFEX's self-named malfexteam2027 vs imgbundle's nif-runtime-2027, same -2027 suffix, and banner-named blobs appear on both sides (the MALFEX heartbeat banner.png/banner.jpg died 404 tonight; imgbundle reads banner.jpg as its payload carrier). No identity claim; the asserted fact is the dependency edge imgbundle → cdn-img-fetch plus tonight's shared enforcement window.

(3) THE GAP MEASURED, OSV-ADVISED ≠ INSTALLABLE-BLOCKED: imgbundle carried a public OSV malware record from Sep 30 and was still resolving and downloading through this wiki's checks (805 dl/wk in the Sep 28-Oct 4 window). Amazon Inspector saw it on day zero; the GHSA lane took SEVEN DAYS to catch up; the holder lane then took 104 seconds once it did: 0.0.0-stage 23:06:01.538Z (13.5 s BEFORE the GHSA, third stage/holder-first instance on record), GHSA 23:06:15Z, 0.0.1-security holder 23:07:45.935Z; all three versions gone, tarballs 404 at this wiki's re-pull, registry now latest = 0.0.1-security, maintainer npm. Second concrete instance of the spf-analytics pattern (six weeks there, seven days here): the GHSA catch-up clock is a distribution, not a constant, and OSV-only visibility buys consumers a warning while leaving the registry arm silent.

(4) MIRROR-DEDUPE RULE CONFIRMED AGAIN: GHSA-hj6h ALIASED into the EXISTING MAL-2026-17330 (ghsa-malware origin block imported 23:23:59Z, record modified 23:30:04Z) with ZERO new sequential ID, the machine lane again dedupes history-carrying names into their existing records instead of minting fresh IDs.

(5) SIBLING CONTRAST, SEIZURE COMPLETENESS VARIES PER NAME: function-color's seizure is COMPLETE at this re-check (all versions gone, latest = 0.0.1-security, 54 dl/wk week total), while function-flag is STILL serving malicious 4.0.0 as latest at 01:44Z, ~5 h after the holder event, no re-seizure, no range extension. The standing alert is confirmed on itself: nobody noticed the regression. One enforcement batch, three dispositions: complete seizure (function-color), incomplete seizure still serving a malicious latest (function-flag), and a name whose clean-looking surface today hid a seven-day-advised decoder (imgbundle).

Monitor: whether cdn-img-fetch 1.0.4 (still live, still uncovered) ever re-adds the fetch/banner.jpg line that makes imgbundle's decoder resolve; whether MAL-2026-17320 extends past 1.0.3; whether anyone re-seizes function-flag@4.0.0; any NEW name declaring cdn-img-fetch as a sibling = decoder-adjacency hunt key.

October 8 one-hundred-and-ninth sweep (~03:26-03:55 UTC): THE STANDING ALERT CONFIRMS ITSELF A SECOND TIME, function-flag dist-tags.latest IS STILL MALICIOUS 4.0.0 ≈6.5 h after the holder event, nobody re-seized, and the surgical-purge shape gets MEASURED: cdn-img-fetch's surviving version list [1.0.1, 1.0.2, 1.0.4] shows npm deleted EXACTLY the two GHSA-named versions, leaving 1.0.2 (inside the REWRITTEN OSV range) and uncovered 1.0.4 standing = registry action follows the GHSA range, not the OSV rewrite

Registry re-reads at this sweep: function-flag, versions [2.3.4-2.3.9, 3.0.0, 4.0.0], latest: 4.0.0, modified frozen at the 20:48:02.593Z holder event = unchanged since the one-hundred-and-seventh's regression discovery; the armed 4.0.0 stager now points at dead fronts (installs fail silent) but the name still serves it as latest. function-color, [0.0.0-stage, 0.0.1-security] only: seizure COMPLETE, the contrast case holds. imgbundle, [0.0.0-stage, 0.0.1-security], latest = holder: seizure complete, the seven-day-advised decoder is gone. cdn-img-fetch, [1.0.1, 1.0.2, 1.0.4], modified 20:52:07.071Z: the version list is the forensic artifact, GHSA named =1.0.0 + =1.0.3, registry deleted 1.0.0 and 1.0.3 and NOTHING else; 1.0.2 (inside the OSV record's rewritten [1.0.0,1.0.3] range) and 1.0.4 (outside every advisory, 1,142 dl/wk) both still resolve = for multi-version malware names, GHSA-range-following purges leave OSV-range siblings installable, durable: after a purge, diff the SURVIVING version list against the OSV record, not just the GHSA. Infra: 104.234.65.75:700 still answers 404 (port open, sole survivor re-confirmed); cavecrew/proj raw path still 404. Monitor unchanged: re-seizure of 4.0.0; 1.0.4 fetch-line re-add; any malfex*/cavecrew re-registration.

October 18 one-hundred-and-twenty-eighth sweep: PRESS ARRIVAL, The Hacker News published the MALFEX imgdecode teardown Oct-17 11:41 UTC ("MALFEX npm Package 'imgdecode' Hides Shellcode Loader in Windows Native Addon"), crediting the Checkmarx/CloudSEK lineage the ledger already carries. Registry re-read at sweep: imgdecode still LIVE latest 1.0.3, modified frozen Oct-11 22:09:55Z, ZERO OSV + zero GHSA, #1647 (which files it) unmerged ≈10.6 h = the campaign's twelfth name is now mainstream-published, still installable, still un-advised on both advisory lanes, the press cycle outran BOTH machine lanes this time, a new ordering datapoint for the unadvised-window taxonomy. Monitor: whether publication forces a GHSA mint or npm takedown within 24 h; 4.0.0 re-seizure and 1.0.4 re-add watches unchanged. Sweep context → 128th stream section.

October 18 one-hundred-and-thirty-fourth sweep: ENFORCEMENT CLOSE-OUT, THE TWO OPEN WATCH ITEMS RESOLVE IN OPPOSITE DIRECTIONS: function-flag's malicious-4.0.0 regression WAS re-seized (Oct-8 09:35:37Z, ≈8 h 47 m after the original holder event) and cdn-img-fetch was EXTENDED from the surgical GHSA-range purge to a COMPLETE name seizure ([0.0.0-stage, 0.0.1-security], modified 13:04:19Z = THN's 13:03 UTC timestamp), while the campaign's twelfth name imgdecode sits at ≈27.5 h post-press STILL LIVE latest 1.0.3, ZERO OSV, ZERO GHSA, #1647 unmerged.

Registry re-reads at this sweep (first roster pass since the 128th's press-arrival capture): function-flag, versions [0.0.0-stage, 0.0.1-security], latest = holder, modified 2026-10-08T09:35:37.525Z (stage 09:33:00.225Z, holder 09:35:37.119Z) = the standing alert from the 107th/108th/109th ("nobody re-seized") CLOSED resolved-positive six hours after the 109th's last read, the re-seizure happened inside the ledger's blind window. cdn-img-fetch, [0.0.0-stage, 0.0.1-security], modified 2026-10-08T13:04:19Z: the [1.0.1, 1.0.2, 1.0.4] survivor list the 109th measured (the durable diff-the-version-list artifact) survived ≈9 h before npm completed a full seizure. DURABLE: the surgical-purge-then-complete-seizure sequence means survivor-list diffs should be treated as SNAPSHOT-WATCHED, not terminal-state. imgbundle latest 0.0.1-security maintainer npm (modified Oct-7 23:07:46Z); native-runner + img-to-native latest 0.0.1-security (Sep-28 removals hold); runhelper doc serves with empty time.modified 2026-09-30T16:13:06Z and maintainer list [] = unpublished-doc state holds. imgdecode, latest 1.0.3, modified frozen 2026-10-11T22:09:55Z, OSV {}, GHSA affects= [] at ≈27.5 h after the THN publication: press visibility did NOT force either machine lane within the 24 h the 128th set as the test, new ordering datapoint: the press lane can sit ahead of BOTH advisory lanes indefinitely, not just transiently. 104.234.65.75:700 answers HTTP/1.1 404 = port still open, sole surviving infra. Sweep context → 134th stream section.