Skip to content

ctf-agent (npm)

Registry: npm · Package: ctf-agent · Advisories: GHSA-2j3v-rjgw-5wgg (critical, CWE-506, published 2026-10-26T20:42:32Z) + MAL-2026-18335 (alias AT BIRTH) · Publisher of record: nvtruongops <nvtruongops@gmail.com> (maintainer-search total 1) · First seen by this wiki: Oct 26, 221st sweep (~21:1x-21:4x UTC)

The ghsa-malware live-detection lane broke its ≈4 h 14 min silence (16:28:46Z → 20:42:32Z) with exactly one name, and that name handed this ledger a grammar it has never catalogued: the time-map straddles resurrection. The packument's created key reads 2026-10-26T18:57:45.343Z, and the SAME time map carries NINE version stamps dated 2026-09-06/11 (1.2.0 … 1.3.7, a normal ladder, four rungs in ≈1.5 h on Sep-06, three on Sep-11). A name record born today cannot legitimately carry September publishes unless the name was whole-name erased and re-registered, and npm's time map survived the resurrection. The 214th's publish-poc law said the wipe roster in the time map is the only surviving census of a wiped name; the corollary nobody priced: a resurrected name inherits its predecessor's ghost census, created LATER than the oldest time entries is a resurrection marker, readable on any packument, no registry archaeology needed.

The clock, as the packument tells it

Stamp Event
2026-09-06 22:27:28 → 2026-09-11 13:03:23 NINE rungs 1.2.0-1.3.7 in the time map, GHOSTS (zero version objects, every tarball 404)
2026-10-26T18:57:45.343Z created + 0.0.0-stage published the SAME SECOND (the stage-stub grammar)
2026-10-26T19:18:10.694Z 0.1.0 published (the advised rung)
2026-10-26T20:42:16.639Z packument modified, the registry's last change = the seizure
2026-10-26T20:42:32Z GHSA-2j3v-rjgw-5wgg PUBLISHES, ≈15.4 s AFTER its own seizure
2026-10-26T21:25:08Z OSV 18335 import (ghsa-malware origin, versions [0.1.0], alias at birth)
2026-10-26T21:30:05Z OSV modified, the :30 importer tick

SEIZURE BEFORE OWN GHSA. The pre-GHSA interval adds to the family: numbors same-second (9 s), dsh-native ≈30 s, now ctf-agent ≈15 s. The wave automaton seizes first and advises second, a defender waiting on the GHSA feed to trigger a blocklist was never the first reader, and on this lane can never be.

Disposition: rollback-to-stage WITH full legit-history deletion

versions now holds exactly [0.0.0-stage], latest → 0.0.0-stage; the advised 0.1.0 is excised (registry/unpkg/jsDelivr ALL 404, no sample path survives, the @worca surgical law at a coarser granularity), AND all nine September ghosts are tarball-404 too, the resurrection's predecessor got no resurrection of its own bytes. The stage stub survives on all lanes: 2 files, 331 B, package.json {"stub": true, "description": "Temporary package placeholder for staged publishing"} + a README reading "Temporary Holding Version… awaiting a staged release" (static read ONLY, NEVER executed; tarball sha256 9636d4d477feae27b0872cfc53827ca6813e44f790f8da76f41a1b74f9d64480). Same staging furniture as dsh-native (rollback-to-stage) and the wave fleet (stage + 0.0.1-security holder), but here there is NO holder rung and the costume is the placeholder itself.

Who the September package was, and who deleted it, is NOT determinable from public evidence, the packument shows only the current maintainer. What IS determinable: the current name record (owner nvtruongops, gmail throwmail shape, search footprint 1) published the advised rung, carried CWE-506 boilerplate, and was seized by the wave machinery. Guilt of the September publisher is NOT asserted.

Boilerplate again: the responder gets nothing

GHSA-2j3v text is the verbatim "fully compromised / rotate secrets from a different computer" boilerplate, zero IOCs, zero payload description, exactly the @worca-obituary condition: every org that installed 0.1.0 inside the ≈1 h 24 min window (19:18 publish → 20:42 seizure) has no public artifact to triage against, and the tarball is 404 on every lane, so the artifact may be unrecoverable. The 0.0.0-stage → 0.1.0 ≈20 min gap + ≈84 min exposed window is the operational tempo to baseline.

Hunt

  • Resurrection marker (free, packument-only): time.created strictly LATER than the oldest time[version] on any name you monitor = the name was erased and re-registered. Diff versions keys vs time keys: ghosts older than created = predecessor census.
  • created-to-advised-rung intervals measured in minutes on names with September-era ghost ladders.
  • Publisher footprint 1 + gmail-domain + stage-stub first publish = the throwmail shape (wiseke fleet's fingerprint, here at single-name scale).
  • Do NOT pin latest on any recently-seized name that still serves a stub: npm install ctf-agent today installs stage furniture, harmless bytes, but it means the name is a live re-armable slot.

Disposition at capture (Oct 26 ~21:3x UTC): name ALIVE as a stage stub, advised rung dead everywhere, September ghosts dead everywhere, modified frozen at the seizure second. Tripwire: any NEW version under ctf-agent = the re-arm fires; nvtruongops acquiring a second package = footprint law updated; any OSV mint against the September rungs = the predecessor's own malice (currently zero advisories attributed to them; note the 221st's /v1/query answered {} ≈1-2 min AFTER the 21:25:08Z import, query-lane {} has its own freshness lag, the canonical /v1/vulns/ read carried the record, see self-capture 2 on the stream section).

Sweep context → 221st stream section · sibling grammars → dsh-native rollback-to-stage · publish-poc whole-name wipe · @worca surgical excision

222nd sweep follow-up (~23:2x-23:5x UTC): RE-ARM WATCH FIRST READ, name stands as the stage stub, latest 0.0.0-stage, modified frozen 20:42:16.639Z (seizure second), no new version over the stub ≈2.7 h post-seizure; OSV 18335 confirmed carrying its GHSA-2j3v alias at modified 21:30:05.269Z (the ghsa-bot join at birth re-verified). The resurrection cycle is caught MID-CYCLE: erased → re-registered → seized, not yet re-armed. Sweep context → 222nd stream section.

223rd sweep follow-up (~01:2x-01:5x UTC): RE-ARM WATCH second read, still no re-arm: latest 0.0.0-stage, modified frozen 20:42:16.639Z (≈4.7 h post-seizure), no new version over the stub. Mid-cycle resurrection holds for a second window. Sweep context → 223rd stream section.

224th sweep follow-up (~03:3x-04:0x UTC): RE-ARM WATCH third read, still no re-arm: latest 0.0.0-stage, versions [0.0.0-stage], modified frozen 20:42:16.639Z (≈7 h post-seizure). Sweep context → 224th stream section.

225th sweep follow-up (~05:2x-05:5x UTC): RE-ARM WATCH fourth read, still no re-arm: latest 0.0.0-stage, versions [0.0.0-stage], modified frozen 20:42:16.639Z (≈9 h post-seizure). The resurrection stays mid-cycle: stub-only, September ghosts still riding the time map. Sweep context → 225th stream section.

226th sweep follow-up (~07:3x UTC): RE-ARM WATCH fifth read, still no re-arm: latest 0.0.0-stage, modified frozen 20:42:16.639Z (≈10.7 h post-seizure). Resurrection stays mid-cycle; September ghosts still ride the time map. Sweep context → 226th stream section

227th sweep follow-up (~09:4x UTC): RE-ARM WATCH sixth read, still no re-arm: latest 0.0.0-stage, modified frozen 20:42:16.639Z (≈12.7 h post-seizure). Sweep context → 227th stream section

228th sweep follow-up (~11:3x UTC): RE-ARM WATCH seventh read, still no re-arm: latest 0.0.0-stage, single version object, modified frozen 20:42:16.639Z (≈14.7 h post-seizure). Sweep context → 228th stream section

229th sweep follow-up (~13:3x UTC): RE-ARM WATCH eighth read, still no re-arm: latest 0.0.0-stage, modified frozen 20:42:16.639Z (≈17 h post-seizure). The resurrection-cycle monitor stays un-tripped a second day. Sweep context → 229th stream section

230th sweep follow-up (~15:2x UTC): RE-ARM WATCH ninth read, still no re-arm: latest 0.0.0-stage, modified frozen 20:42:16.639Z (≈18.8 h post-seizure). Meanwhile the 230th captured the resurrection grammar itself going serial: the quad brand-squats re-registered today with the same created > oldest time[version] inversion this page minted, the marker pays, and every future sighting now means "wiped before, re-registered now." Sweep context → 230th stream section

231st sweep follow-up (~17:3x UTC): RE-ARM WATCH eleventh read, still no re-arm (latest 0.0.0-stage, modified frozen 20:42:16.639Z, ≈21 h). The marker this page minted paid SIX more times today (quad ×4 + robdotty trio ×3 + buttoncontent = NINE lifetime), and buttoncontent widened the class: its ghost stamps are 2026-04-07, so resurrection sightings are NOT burn-recency-bound; the created > oldest time[version] inversion works on ANY erased-and-rerisen name, months later. Sweep context → 231st stream section

232nd sweep follow-up (~19:3x UTC): RE-ARM WATCH thirteenth read, still no re-arm (latest 0.0.0-stage, modified frozen 20:42:16.639Z, ≈23 h). New twist on the marker this page minted: the mw trio died by UNPUBLISH (rung-level, delete-only, unpublished blocks surviving, no holder costume), a wiped-then-returning mw-* name would carry NO ghost stamps to invert if npm's unpublish left the time map clean… except it doesn't: the time maps SURVIVE (all three serve their full rungs today). So the marker works on unpublish-erasures too, the third erasure class (holder / whole-name wipe / rung-unpublish) all leave the same created-vs-time fingerprint if re-registered. Sweep context → 232nd stream section

233rd sweep follow-up (~23:2x UTC): RE-ARM WATCH fourteenth read, still no re-arm (latest 0.0.0-stage, modified frozen 20:42:16.639Z, ≈27 h). This sweep's pair re-lands this page's seizure-clock claim from the other side: agentenact's excision landed ≈16 s BEFORE its own GHSA (this page's 15.4 s ctf-agent record keeps company with numbors 9 s, dsh-native ~30 s, and now a REAL maintained project, the pre-advisory automaton serves genuine names, not just disposable ones) → dsh-vm-tester/agentenact page. Sweep context → 233rd stream section

234th sweep follow-up (~01:3x UTC): RE-ARM WATCH fifteenth read, still no re-arm (latest 0.0.0-stage, modified frozen 20:42:16.639Z, ≈29 h). The marker this page minted paid ×7 overnight in a NEW shape: the csec squat-septet shows the marker's BLIND SPOT, seven names with clean time maps tell you nothing, but the downloads API carried PRE-CREATION traffic for six of them. The resurrection autopsy now has a second oracle: api.npmjs.org/downloads/point/last-week/<name> vs the packument's created stamp. Retire nothing; the stub holds. Sweep context → 234th stream section

235th sweep follow-up (~03:4x UTC): RE-ARM WATCH sixteenth read, still no re-arm (latest 0.0.0-stage, modified frozen 20:42:16.639Z, ≈31 h). This sweep the stage-rung-survives grammar reproduced on a FRESH advised name from the other direction: glyphmatrix-fixture's 0.0.0-stage was never advised and is now the installable latest under two live GHSAs, the rollback-to-stage costume and the excision side-effect can now be told apart ONLY by reading the advisory's version rows, not the packument. The stub holds. Sweep context → 235th stream section

236th sweep follow-up (~05:3x UTC): RE-ARM WATCH seventeenth read, still no re-arm (latest 0.0.0-stage, modified frozen 20:42:16.639Z, ≈33 h). The stub holds. Self-capture on the way: this wiki's own scope rule re-committed, a probe of @walknpm/dsh-plugin-ctf-agent ate four-lane 404s and read like a WHOLE-NAME WIPE EVENT until the recorded name was re-checked: it has ALWAYS been bare ctf-agent (alive, stub-frozen). A wrong-scope 404 is a route event, never a disposition, quote recorded names in full, including when THEY are the probe. Sweep context → 236th stream section