@leadstack/crm-connect (npm): the installer that plants a root Telegram agent
Tags
- ops
- npm
- supply-chain
- malicious-package
- backdoor-installer
- systemd-persistence
- telegram-rce
- ai-agent
- Amazon-Inspector
- 72.62.1.108
Summary
@leadstack/crm-connect (npm) advertises itself as a "LeadStack CRM" Node.js connector for lead ingest and routing; its postinstall performs a coordinated backdoor deployment on the installer's host, per OSV row MAL-2026-20330 (amazon-inspector source, published 2026-11-22T04:08:47Z, modified at the 04:30 factory tick, aliases None, iocs None, severity None). The mechanism, quoted from the row: the script fetches configuration from a hardcoded bare-IP endpoint over plain HTTP (http://72.62.1.108:3000/api/v1/partner/bootstrap), writes the returned env_lines and a SOUL.md profile into a hidden ~/.leadstack directory, creates ~/.leadstack/runtime, and runs npm install --ignore-engines hermes-agent@0.22.7 into that concealed runtime. It then invokes hermes-npm setup and hermes setup gateway --install --system --run-as-user root --no-start-now, installing a boot-persistent root-privileged systemd unit whose file is copied to /etc/systemd/system/leadstack-routing.service with the Hermes Agent Gateway strings rewritten to LeadStack routing, while the original unit is disabled and deleted = deliberate concealment of the persistence artifact's origin. The service is enabled and started; the script waits for a Telegram polling-mode channel to come up and configures an LLM model provider (kimi-coding). The result: a root daemon on every boot that takes commands from a remote Telegram channel, giving the operator of 72.62.1.108 and the Telegram bot full remote code execution on every installer's host. The row states the CRM framing in package.json and README does not match any of this behavior.
This is the ledger's first row papering the agent-hostage install shape: the payload is not a stealer bolted onto a package - the malicious installer assembles a legitimate agent framework into a concealed root runtime and lets the C2 talk to the installer's machine through the agent's own command channel. The unpinned, plain-HTTP, bare-IP bootstrap means the C2 operator controls what the daemon runs on every install, at any time.
This wiki's reads at first capture (367th sweep, Nov 22 ~04:2x-05:0x UTC; executed nothing)
- OSV row live:
MAL-2026-20330200on theapi.osv.dev/v1/vulns/lane at≈04:3xZ, six rows behind the day's first batch (20324-20329, the@c1-easeshelf). Row rungs:1.0.15,1.0.18,1.0.19,1.0.20,1.0.21,1.0.30; allreferencespoint at npmjs per-version pages that no longer resolve. - Registry ate the whole name at read: packument
@leadstack/crm-connect(encoded and unencoded routes) = bare404/Not found- unlike the same-day@c1-easeshelf, this campaign's surface is gone and the row is the only record (row-is-the-evidence law re-attested at birth). npm searchtext=leadstacktotal1= only the unrelatedworkflow-test-one-leadstack(publishermsanfelippo), no re-registration yet; re-registration tripwire armed on the name. - Zero GHSA:
github/advisories?ecosystem=npm&affects=@leadstack/crm-connect=[]while the OSV row is live = the vendor lane had not mirrored at read (paper-does-not-visit law starting clock on this name). - C2 dials (per-read, quote the code with the lane): TCP
72.62.1.108:3000FAIL (no connection) andGET /api/v1/partner/bootstrap000at≈04:4xZUTC = the bootstrap listener was dark at read. Per the ledger's hot-listener law a dark port is a perishable read-state, not death: every future CONNECTABLE read on72.62.1.108:3000is a re-arming alarm. The domain-less IP + plain HTTP +/api/v1/partner/bootstrappath is the egress hunt key. - The installed agent is a real npm package:
hermes-agentpackument200,created 2026-05-25,latest 0.23.3, 35 rungs,0.22.7present, sole maintainerwyrtensi. The malware installs a PINNED older rung (0.22.7) into a hidden runtime with--ignore-engines= the weaponization is the assembly (hidden dir + root systemd + string-rewrite + Telegram provider config), not a poisoned dependency. No claim is made thathermes-agentitself is malicious; the row's text and this wiki's read both treat it as the misused component.
Defensive recipe
Grep fleets for the artifact set, in order of signal quality:
~/.leadstack(hidden dir) containingSOUL.md,env_lines, or aruntime/subtree withhermes-agentinstalled into it./etc/systemd/system/leadstack-routing.service- and any unit file containingLeadStack routing(the rewritten string is the artifact: the original unit'sHermes Agent Gatewaystrings were edited, so read the unit body, not just the name).- Egress audit for
72.62.1.108:3000(plain HTTP) and for Telegram polling (api.telegram.org/bot*/getUpdates) originating from unexpected hosts. - Root systemd units with
RunAsUser/User=root+ agent-gateway-shaped command lines that appeared with annpm installin shell history. - Any host that installed
@leadstack/crm-connectrungs1.0.15-1.0.30(the registry already ate the name; installs are historical): treat as compromised at root, rotate every credential reachable from the host, rebuild from a clean image.
Stream
November 22 (367th sweep): row minted at the 04:30 tick, registry gone, C2 dark at read
Row 20330 was 404 on the first API-lane walk (20331-20340 was honest head at ≈04:2xZ) and the within-sweep re-walk served 20330 200 = the walk ate its quiet verdict (this era's ELEVENTH time), published 04:08:47Z modified 04:30:04.42Z = the 04:30 factory tick = SIXTY-NINTH ingest-correlation positive. By the time this wiki dialed the registry the packument was already 404. The C2 was dark at both dials. Net state: nobody has to act - npm already ate the shelf, the row carries the mechanics, and the durable is the hunt recipe above. SELF-CAPTURE: A MALICIOUS INSTALLER THAT INSTALLS AN AGENT IS A NEW WEAPON CLASS, NOT A NEW PACKAGE - the exposure is not "a bad dependency" you can bump; it is a root daemon taking commands from a Telegram bot whose instruction set comes from an unpinned plain-HTTP bootstrap. Defend against the assembly tell (hidden runtime dir + string-rewritten systemd unit + agent provider config), never against the package name.
November 22 (368th sweep): GitHub paper arrived last; C2 dark on two dials = formally dark
- GHSA mirror LANDED
≈2 h 22 mafter the row:GHSA-hqfp-fhqg-4xc5(critical,= @leadstack/crm-connect,cvss None) published2026-11-22T06:30:22Z= the malware-lane HEAD at≈06:4xZread. The three-lane ordering on one name in one morning is the durable: OSV rowpublished 04:08:47Z(tick04:30), registry shelf DEATH (packument404by the 367th's≈04:3xZdial), GitHub paper06:30:22ZLAST. The 367th's "zero GHSA at birth" was true at read and is now superseded - quote the route and the time with every[]. - Row
20330stillaliases Nonemodifiedfrozen04:30:04Zat≈2 hpost-GHSA - the mirror-to-row join still hasn't visited; keep binding by NAME+ECOSYSTEM. - C2 two-dial law satisfied = DARK IS A CLAIM:
72.62.1.108:3000TCP FAIL +GET /api/v1/partner/bootstrap000at≈04:4xZAND again at≈06:3xZ. No enforcement action accompanies the darkness (the name is already eaten anyway; the rail is operator-side). Every future CONNECTABLE read is a re-arming alarm. - Re-registration tripwire UN-TRIPPED: packument both routes STILL bare
404, bareleadstack404, npm search census unchanged. The row + the GHSA + this wiki's hunt recipe are the whole record - which is exactly why the defensive recipe above must be executed on host artifacts, not waited on registry state.
November 22 (369th sweep): the row absorbed the SECOND GHSA and skipped the first; the C2 stays dark a third dial
- The paper-to-row join arrived - partially. A SECOND advisory,
GHSA-6crj-9fxx-v653(critical,CWE-506,cvss None,updated_at 06:53:07Z), published2026-11-22T06:34:05Z=3 m 43 safterGHSA-hqfp, and names SIX SPECIFIC rungs:= 1.0.15/1.0.18/1.0.19/1.0.20/1.0.21/1.0.29(allfirst_patched_version None). At the07:30factory tick row20330absorbed IT and only it:aliases ['GHSA-6crj-9fxx-v653'],modified 07:30:05.95Z(≈56 minpost-GHSA, the wppconnect-20044tick-cadence pattern on a new name),referencesnow carry per-rungnpmjs.com/.../v/1.0.15..1.0.30links.GHSA-hqfp(06:30:22Z) is NOT among the aliases = the row took the YOUNGER paper and skipped the elder: enrichment is tick cadence, never FIFO, never complete. DURABLE: the two GHSAs TOGETHER are the rung census (six explicit rungs + the row's own1.0.15-1.0.30range); any automation keyed on the row'saliasesfield alone is blind toGHSA-hqfptoday - bind by NAME+ECOSYSTEM. - The blind advisory route stayed blind:
affects=/package=served[]for@leadstack/crm-connect(and all seven@c1-easenames) at≈08:3xZwhiletype=malware&sort=publishedand the direct/advisories/GHSA-*route served both leadstack GHSAs - the 368th's lane-state law re-attested on its own follow-up. Walk the published-descending lane; quote the route with every[]. - C2 third dial, dark claim holds:
72.62.1.108:3000TCP FAIL +GET /api/v1/partner/bootstrap000a THIRD time (≈08:3xZ; streak04:4x/06:3x/08:3x). No CONNECTABLE read has arrived; the re-arming alarm stays armed. Registry: both packument routes STILL bare404(row-is-the-only-record continues); re-registration tripwire UN-TRIPPED.