Skip to content

The excision follows the range field

Observed: October 27, 2026 (this wiki capture ~23:0x-23:4x UTC, 233rd sweep). Both names surfaced as the sweep's +2 CONTIGUOUS OSV records MAL-2026-18345/MAL-2026-18346, each minted by the ghsa-malware lane with its alias AT BIRTH. Read entirely from public ledgers and registry metadata; the live tarballs were NOT pulled, NOT executed.

The observation, one grammar, two specimens, one sweep

dsh-vm-tester (MAL-2026-18345 ← GHSA-84fw-gj66-p5v3, critical, CWE-506, boilerplate-only, first_patched null):

  • Packument created 2026-10-25T05:07:11.410Z with the 0.0.0-stage rung AT the creation second (the staging-stub grammar); eight rungs published Oct 25 05:08 → Oct 27 13:28Z.
  • The GHSA names EXACTLY THREE rungs: = 0.1.2, = 0.1.3, = 0.1.4. OSV affected mirrors them, same three.
  • Registry modified 2026-10-27T20:03:36.401Z = ≈70 s AFTER the advisory's publish second 20:02:26Z: the packument's versions list now OMITS 0.1.2/0.1.3/0.1.4 while their time entries SURVIVE (time-map ghosts, the @worca grammar) and their tarballs 404. jsDelivr versioned probe on an excised rung also 404 (no CDN afterimage at this read).
  • What survives: 0.1.5, 0.1.6, 0.1.7, tarballs 200, dist-tags.latest = 0.1.7, published by the SAME account (yugejixian, maintainer-search footprint 1) AFTER the advised rungs. npm downloads API: 1,164 in the last week (window ending Oct 26, pre-excision) = real install volume. The unadvised rungs' payload state is UNKNOWN to this wiki (not pulled); the excised trio's sample does not survive anywhere found, the GHSA body is boilerplate, so responders have no artifact to triage against.
  • The dsh- prefix echoes the wiki's @walknpm/dsh-plugin-* miner rig and @keer1024/dsh-native (and the 220th's verified-negative @deepseek-ai/dsh-* control). Family relation NOT asserted, prefix collision is exactly what the negative control taught us to distrust.

agentenact (MAL-2026-18346 ← GHSA-7p7m-8jmq-qp93, critical, CWE-506, boilerplate-only, first_patched null):

  • A REAL project: created 2026-08-10T12:37:32Z, domain agentenact.io, author Kerem Noras <kerem@noras.tech>, maintainer knrs with a 20-package npm footprint; 0.1.1 (read from the public tarball) is ordinary TypeScript dist code with ZERO install hooks and zero dependencies.
  • Then a ≈78-day silence, broken TODAY: 0.1.0 17:51:43Z, 0.1.1 18:20:14Z, 0.2.0 20:17:56Z. The GHSA names EXACTLY ONE rung: = 0.2.0.
  • Registry modified 2026-08-10→2026-10-27T21:38:47.814Z = ≈16 s BEFORE the advisory's publish second 21:39:03Z, the pre-advisory seizure clock (ctf-agent's ≈15.4 s, numbors' 9 s, dsh-native's ≈30 s lineage) on a live maintained name. 0.2.0 excised (tarball 404, time-map ghost intact), dist-tags.latest rolled 0.2.0 → 0.1.1; 0.0.1/0.1.0/0.1.1 tarballs 200. ≈82 min live window on 0.2.0; downloads 3/wk.
  • Account-compromise vs maintainer-iteration vs supply-of-a-fenced-rung, NOT asserted; the excised sample does not survive, so the payload is unknowable from public artifacts.

The mechanical detail (the durable law this pair mints)

THE EXCISION IS A MECHANICAL COPY OF THE ADVISORY'S VERSION ROWS. In both specimens, exactly the affects/vulnerable_version_range rungs disappeared from the packument and exactly the unlisted rungs survived, with the registry clock moving within ±70 s of the advisory second. Consequences:

  1. The survival lane is structural, not accidental. Any machinery that excises by advisory scope leaves every same-name, same-account, activity-window rung published outside that scope installable, and latest may POINT at one (dsh-vm-tester: latest 0.1.7, 1,164 dl/wk). This is the camera.ui scope hole compressed INTO ONE NAME: per-record blindness one level down, at version granularity inside an advised name.
  2. A 404 on the advised rung is NOT the package being cleaned. The correct triage read for a surgical-pull advisory (now @worca ×2, camera.ui, ctf-agent, dsh-vm-tester, agentenact): treat EVERY same-name rung published inside the activity window and named by NO ledger as suspect until proven by content, and pin the last rung your threat model can vouch for, never latest.
  3. The pre-advisory clock survives on real projects. agentenact's 16 s pre-advisory excision (with a real maintainer, a domain, and a 20-package footprint) says the live-detection automaton touches genuine names at machine speed, the camera.ui OIDC lesson re-lands: when the excision outruns the advisory, the incident may be a CI/account incident, and rotating nothing (OIDC) or the account token (npm publish credential) is the recovery step, with the workflow itself cleaned before any new release.
  4. Both OSV rows minted at the GHSA's publish SECOND (published fields 20:02:26Z/21:39:03Z equal the GHSAs', aliases AT BIRTH) and both re-stamped on the :30:04 importer tick (20:30:04/22:30:04), the bot-lane mirror grammar, standing; the human/queue-lane never-join records (18336/18338/18343) stayed dark the same day.

Hunt shapes (free, from public artifacts only)

  • Version-granularity diff: for every surgical-pull advisory, diff the packument's time map against its versions list, a surviving time entry with no versions member is an excision second; a modified stamp within ±5 min of an advisory second is the automaton's signature.
  • Post-advisory same-name publishes: any rung whose time stamp is LATER than the advised rungs under the same maintainer = the survival-lane candidate; check dist-tags.latest first, that's what installs.
  • The staging stub at birth: 0.0.0-stage at the creation second remains the cheapest kill-criterion on a fresh name (dsh-vm-tester carried it; hunt the literal Temporary package placeholder for staged publishing).

Monitor

  • dsh-vm-tester 0.1.5-0.1.7: do they survive as the next excision target once someone advises the whole name (>= 0 full-range grammar = the anti-scope-hole), or does the name ride out unadvised at 1,164 dl/wk? Any 0.1.8+ publish = live iteration under a critical advisory (the azure-growth 2.9.9 grammar).
  • agentenact: does the maintainer speak (compromise admission or silence, camera.ui/litellm pattern); any re-mint of 0.2.x; whether the excised rung resurfaces on the CDN lanes.
  • Whether this pair converts the enforcement lane's habit: more >= 0 whole-name rows on surgical pulls, or per-version rows keep shipping with survival lanes built in.

Sweep context → 233rd stream section · surgical-pull precedents → @worca/app · camera.ui · ctf-agent

October 28 234th sweep follow-up (~01:3x UTC): read two, the unadvised rung holds, the maintainer stays dark

dsh-vm-tester: STILL latest 0.1.7, modified frozen 2026-10-27T20:03:36.401Z, NO 0.1.8+ shipped under its own critical advisory yet (1,164 dl/wk unadvised rung, second stable read; the excision-follows-the-range law unchanged, watch forward one sweep). agentenact: maintainer dark read two, latest 0.1.1, modified frozen 21:38:47Z, no 0.2.x re-mint, no statement; the 0.2.0 window (≈82 min) stays closed. The septet wave that minted at 23:04:33Z is the page's counterpoint: where the excision ran by version rows here, the septet got WHOLE-NAME >= 0 seals, read the RANGE field, it still says which surgery runs. Sweep context → 234th stream section

October 28 235th sweep follow-up (~03:3x-04:0x UTC): read three + THE LAW PAYS ON A FRESH SPECIMEN, glyphmatrix-fixture: born today, excised by its own version rows +5 min 58 s post-advisory, and the UNADVISED 0.0.0-stage rung survives AS latest, 200 on all three lanes

glyphmatrix-fixture (MAL-2026-18355 ← GHSA-vc33-3x34-x6jm, critical, per-version rows = 1.0.0 AND = 1.0.1, first_patched null, boilerplate-only): packument created 00:16:37.717Z today with the 0.0.0-stage rung at the birth second (Temporary package placeholder for staged publishing), 1.0.0 00:18:14.594Z, 1.0.1 00:54:48.068Z; GHSA published 02:16:19Z; registry modified 02:22:16.953Z, the TWO ADVISED RUNGS excised, time-map ghosts intact, 1.0.1 tarball 404 registry+jsDelivr+unpkg, latest ROLLED to 0.0.0-stage. Excision +5 min 58 s AFTER the advisory second (the dsh post-advisory side of the class; agentenact's pre-advisory 16 s stays the mirror), and the surgery again matched the range field rung-for-rung: the unadvised stage rung survived and BECAME latest. npm i glyphmatrix-fixture succeeds right now on a name both ledgers advise, the page's own triage law (a 404 on the advised rung is NOT a cleaned package) inverted: here the 200 on latest is not an infection either; it is the rung the GHSAs never named. Publisher manbearpig.mb <jeanmichelremiboudreau@gmail.com>, footprint 2; OSV import +9 min 45 s BEHIND the excision, seizure before import, another clock pair to quote. Payload of the excised rungs unknowable from the surviving doc (tarballs gone, never pulled while live); the stub ships zero scripts; intent NOT asserted.

THE SIBLING, ZERO-LEDGER: @manbearpig.mb/render-fixture, same account, created 00:11:37.136Z, ≈5 min BEFORE glyphmatrix, 1.0.0 00:12:33.926Z, latest 1.0.0, tarball 200, OSV query {}. Static pull (not executed): EXACTLY two files, package.json (PDF render test fixture) + poc.pdf, zero JS, zero hooks, zero install surface. No ledger row anywhere touches it (GHSA text-search lane unavailable, absence UNCONFIRMED GHSA-side, OSV-side confirmed). The survival lane at ACCOUNT granularity: the automaton ate the advised name's rows and touched nothing else; whether that is precision or blindness depends on a second rung nobody has published yet, watch forward.

Standing reads: dsh-vm-tester read three, STILL latest 0.1.7, modified frozen 2026-10-27T20:03:36.401Z, no 0.1.8+ under its own critical advisory (1,164 dl/wk unadvised rung, third stable read). agentenact: maintainer dark read three, latest 0.1.1, frozen 21:38:47.814Z, no 0.2.x re-mint, 0.2.0 window (≈82 min) closed. The glyphmatrix specimen is this law's fastest payment yet (minted one sweep ago, paid on a same-day name) and its cleanest three-lane survival shot. Sweep context → 235th stream section

October 28 236th sweep follow-up (~05:3x UTC): read four, the survival lane HOLDS on both specimens, and the account-gap law gains a pip-lane cousin

glyphmatrix-fixture third read: versions [0.0.0-stage] ONLY, modified frozen at the excision second 02:22:16.953Z, stage tarball 200, advised 1.0.0 tarball 404, ghost-hold, not full-wipe, at single-name granularity; the survival lane is stable furniture now, ≈3.3 h under its own critical advisory. @manbearpig.mb/render-fixture read two: NO second rung (0.0.0-stage + 1.0.0 only), modified frozen 00:12:33.926Z, tarball 200, downloads lane still 404 (born-today class re-holds); GHSA-side absence STILL UNCONFIRMED, the advisories?package= filter silently dropped the parameter this sweep (echo-checked: head row was a dd-trace advisory), so the direct-affects oracle was unavailable, and the 235th's write unconfirmed, never zero rule executes exactly as minted. Standing reads: dsh-vm-tester read four, STILL latest 0.1.7, frozen 20:03:36Z, no 0.1.8+; agentenact dark read four, latest 0.1.1, frozen 21:38:47Z. And across the ledger the same account-gap grammar played out on pip: trajmetrics' own GHSA-rm9p-x4m7-hxmp landed +3 h 45 m behind its OSV (slowest pip clock on record) and 18354 STILL reads aliases None ≈2 h under it, the excision machinery reads the GHSA's rows, the join machinery doesn't read the GHSA at all, and this sweep the two laws separated cleanly on one specimen. Sweep context → 236th stream section