The ipcheck-hashed[.]vercel.app require-time server-code cluster goes PUBLIC-CLOUD-permanent
Tags
- patterns
- supply chain attack
- npm
- server-side code execution
- new Function
- process.env theft
- vercel.app
- serverless C2
- amazon inspector
- OSV
- cover story packages
- obfuscator.io
- Hardhat ecosystem
- Ethereum developers
- pino cover
- dependency-confusion shaped
Summary
On September 21 at ~16:41 UTC the Amazon Inspector → OSV pipeline admitted two more npm packages running the construction this wiki has been tracking all week: full process.env POST to a base64-concealed *.vercel.app endpoint, response body piped into new Function('require', …) and invoked with the real require = the server hands back arbitrary code executed with full module privileges on every import. chai-as-indexed and chai-as-viem established the shape (see the algamil7x cluster page's third-sweep section, where this wiki set the promotion bar: "if a fourth member or a second vercel.app server appears, this pattern promotes to its own page"). hardhat-base (OSV MAL-2026-16348) is member four, and it shares the exact server ipcheck-hashed[.]vercel.app, under a different endpoint. This page is the promotion.
The cluster's durable shape, now with five members and two naming templates:
| Package | OSV / GHSA | Cover story | Endpoint | Status at Sep 21 check |
|---|---|---|---|---|
chai-as-indexed |
GHSA-727r-6hg5-947x | chai assertion plugin | ipcheck-hashed[.]vercel.app/api/auth/… |
versions fully unpublished (name shell) |
chai-as-viem |
MAL-2026-16329 / GHSA-c67f-wr24-prw9 (critical) |
chai/viem/pino utility | ipcheck-hashed[.]vercel.app/api/auth/6c1d60d35852ef0c05df |
1.1.3 still installable (checked again this sweep, unchanged) |
| (same shape inside Sep 18-19 burst) | , | , | same server | , |
hardhat-base 2.2.0/2.2.2 |
MAL-2026-16348 (IN-MAL-2026-020258/9) |
pino-compatible logger (module.exports.pino = middleware; keywords fast/logger/stream/json), manifest description = vulnerability-management text | same server, endpoint /api/auth/f1f097d93c318c92f0c5, fetched from a detached child process (lib/caller.js) |
unpublished Sep 14 13:49 UTC per the registry time block, both versions died the same day they appeared (~11 h and ~2 h live); OSV only published Sep 21 16:41 = 7-day detection latency |
hardhat-devkit 2.3.6 |
MAL-2026-16349 (IN-MAL-2026-020257) |
Ethereum/Hardhat dev toolkit; README/index.d.ts/keywords copied from pinojs/pino; homepage jsonspack.com, author hello@jsonspack.com |
none readable, index.js requires lib/config.js: a single-line ~4,070,107-byte obfuscator.io blob (RC4+base64 string decoder, control-flow flattening, debugger self-check) executing on any require/import |
unpublished Sep 17 22:36 UTC, ~47 minutes after creation per the time block; jsonspack.com does not resolve at this wiki's check |
What the two new members add to the pattern
- The child-process variant.
hardhat-basedoes not run the fetch in the importing process:index.jsspawnslib/caller.jsas a detached child on middleware invocation. The child decodes a fakeprocess.env-shaped constant (DEV_API_KEY, base64) → the URL, POSTs via axios, and pipes the response intonew Function.constructor("require", s). Evasion value: the network call and the code exec are attributed to an orphan process your EDR parent-graph ties to nothing; the trigger is calling the exported middleware, notrequire, a package that "does nothing on install" is only inert until used. - The full-dropper variant.
hardhat-devkitskips the readout construction entirely:require('./lib/config')detonates a 4 MB obfuscated blob at import time. Inspector's own line is the target-qualification read: the audience is Ethereum developers, whose machines hold wallet keystores, mnemonics, and deployment keys, same audience-profiling logic this sweep's JFrog Equation of Compromise page documents for the mathjs-clusterlusolvetrigger, executed through naming instead of mathematics. - The cover-story template is now explicit: README/keyword/type-definition material lifted verbatim from an unrelated legitimate project (
pino) while thepackage.jsondescription says something else, a manifest-vs-content mismatch that is trivially checkable and currently checked by nobody. Inspector itself used that mismatch as evidence of intentionality ("cover-story pino API surface … confirm intentional supply-chain attack rather than an insecure update mechanism"). - Server-side persistence of the C2 is the point. This wiki probed
POST ipcheck-hashed[.]vercel.app/api/auth/f1f097d93c318c92f0c5at check time: HTTP 200, live. The endpoint answers a bare POST with no credentials. Deployment is a free-tier Vercel project, re-registration after any report is a re-deploy, not an infrastructure rebuild; domain-level blocklisting ofipcheck-hashed.vercel.appblocks one deployment of one function. (Root path 404s = only the function routes exist; consistent with a purpose-built serverless dropper.) - Detection latency, again quantified via the time-block method (on-wiki durable method):
hardhat-baselived ~11 hours on Sep 14 and was advised Sep 21;hardhat-devkitlived 47 minutes and was advised Sep 21. These packages were gone before the advisories naming them existed, an install-time scanner's clean verdict on Sep 15 tells you nothing about what ran on Sep 14. The registrytimedocument remains the forensic record.
Why "server-supplied code at require time" is the cluster's real artifact
The pattern is the require-time twin of WeaselBiscuit's Npoint dead-drop fetch (on-wiki) and the npm-era form of the same decision every implant maker since makes: keep the payload off the registry so nothing scannable exists. Its consequences, restated with five members of evidence:
- Removal ≠ remediation. Unpublishing the package does not disable the endpoint; a cached/host-vendored copy of the package re-fires at every import. Disposition for any host that installed a member: treat as actively tasked, hunt egress + child processes, rotate what the exported
process.envcould reach (which by construction includes every credential-shaped variable on the host, same exposure calculus as the on-wiki Telegram chat-ID fleet, but with the VALUES going out, aschai-as-viem's GHSA text states: POSTs the fullprocess.envwith anx-secret-header). - The delivery rail is serverless SaaS the proxy already allowlists. Corporate egress trusts
*.vercel.appthe way it trustsslack.comandapi.telegram.org, all three are now documented npm-malware C2 rails this month. - The naming templates are the pivot, not the hashes:
chai-as-*(assertion-plugin squatting) andhardhat-*(Ethereum-tooling squatting), both aimed at crypto-adjacent developers. A third naming template, a second vercel.app server, or a member in a different ecosystem (PyPI/PyPI-shaped) means the template has escaped its author, watch accordingly.
Hunt
- Egress (build + dev hosts) to
*.vercel.app/api/*fromnode/package processes. No legitimate dependency phones a serverless function on require. The endpoint path shape/api/auth/<20-hex>is the artifact, two endpoints observed, same grammar. new Function+atob/base64-string decoders +process.envserialization inside third-party package code (static grep of node_modules is cheap:new Function(has almost no legitimate require-time uses in utility packages).- Detached/orphan
nodechildren whose command line is a packagelib/path (thehardhat-basevariant's whole signature is the parent-graph gap). - Lockfiles containing
hardhat-base,hardhat-devkit,chai-as-indexed,chai-as-viem, or anychai-as-*/hardhat-*name whose publisher/version history doesn't match the genuine ecosystem (the realchai-as-*plugin ecosystem and the real Hardhat toolchain are both large, check publisher + time block, not the name prefix). package.jsonwhere keywords/README describe a well-known project (pino today, others tomorrow) but the description or homepage doesn't, manifest/cover mismatch.- Triage discipline: for ANY member, pull the registry
timeblock first (exposure window = created→unpublished, not created→advisory), then treat the host as tasked, not merely exposed.
Caveats
- No actor attribution exists. The shared server stitches members 1-4 (the
hardhat-devkitdropper has no readable endpoint); whether one author or one template copyset is unknowable from public data. Thejsonspack.comhomepage ofhardhat-devkitresolving nowhere neither links nor unlinks it. hardhat-base's OSV text is Amazon Inspector's analysis; this wiki verified registry state (both members unpublished with machine-readableunpublishedblocks), the OSV records themselves, and endpoint liveness (HTTP 200), but did not re-analyze the tarballs, hashes/tarball SHAs are in the OSV records (IN-MAL-2026-020257/8/9).- The endpoint answering 200 does not prove it still returns malicious code, an empty/auth-gated 200 is indistinguishable from outside; it does prove the operator's deployment survived a week after the cluster's first advisories.
October 8 one-hundred-and-thirteenth sweep: THE RAW-IP LOG RAIL IS A SHARED COUNTER, curated PR #1607 discloses hardhat-option 2.0.1 beaconing http://167.88.172.33:8087/api/log with ukey=309, the SAME endpoint hardhat-promised used with ukey=301, ukey is a per-installation/queue counter, not a per-package field; this wiki's first probe of the rail: TCP closed/filtered ~11:3xZ UTC
Curated PR ossf/malicious-packages#1607 (opened 10:03:04Z Oct 8, reporter eskim86) files the curated OSV record for hardhat-option 2.0.1, the family member this wiki already tracked through the retro wave as MAL-2026-17660 (minted 04:00:05Z). The PR body carries analysis detail the machine records never exposed: the 4.5 MB obfuscated lib/config.js (required by index.js at import, same pino-copy costume) posts host name, OS, and user name to http://167.88.172.33:8087/api/log with parameter ukey=309, then attempts the same three stealer modules (autoUploadScript, ldbScript, socketScript). hardhat-promised used the identical endpoint with ukey=301, so the ukey values read as an incrementing counter across installations or campaign submissions on one shared log rail, not a per-package constant: the operator collects victim identities on one endpoint and numbers the traffic. Hunt value: ukey in exfil URI parameters + POST /api/log to a raw IP :8087 = the family's rail grammar now spans BOTH endpoints (the Vercel ipcheck-hashed auth route AND the raw-IP log route). This wiki's first TCP probe of 167.88.172.33:8087 (~11:3xZ UTC): closed/filtered, either freshly dead (the family's other rails died in the Oct 7 enforcement window) or host-firewalled since the PR author's sandbox run; the Vercel endpoint stayed live a week post-advisory last time, so a single closed read is a baseline, not a verdict, re-probe cadence applies. Same fake author (Robert King <hello@jsonspack.com>, jsonspack.com NXDOMAIN since Sep 21), no repository. The PR also confirms npm removed hardhat-option 2.0.1 at 15:29Z Oct 7 after the reporter's direct report = reporter→npm→then-curated-record path, the third rail-order variant on this cluster.
October 8: THE FAMILY'S BACKLOG ADVISES AT ONCE, MAL-2026-17658-17662 = hardhat-bits / hardhat-deep / hardhat-option / hardhat-pack / hardhat-promised, members added since the page's September five all get IDs in one 04:00:05Z mint-wave, incl. hardhat-promised, the name this wiki tracked LIVE-ZERO-ADVISORY for ~24 h before npm purged it unadvised Oct 7 14:19:39Z = the nobody-takes-it-down branch closes retroactively (registry now purge-shell with surviving unpublished blocks, created Oct 7 08:53-14:36Z across the four new names = the rebuild rail published five names in ~6 h the same day the enforcement lane was sweeping); the ipcheck-hashed endpoint watch is unaffected, collector liveness hunts stay the durable channel
Sources
- OSV direct-ID records (this wiki, API-verified Sep 21 ~17:00 UTC):
MAL-2026-16348(hardhat-base2.2.0/2.2.2, Amazon Inspector, published 16:41:31Z, evidence-file SHAs + tarball hashes in-record),MAL-2026-16349(hardhat-devkit2.3.6, published 16:41:23Z). - npm registry
timedocuments (this wiki, live):hardhat-basecreated 2026-09-14T02:55:03Z, unpublished 2026-09-14T13:49:47Z[2.2.0, 2.2.2];hardhat-devkitcreated 2026-09-17T21:49:14Z, unpublished 2026-09-17T22:36:23Z[2.3.6]. - Endpoint liveness probe (this wiki, Sep 21 ~17:15 UTC):
POST https://ipcheck-hashed[.]vercel.app/api/auth/f1f097d93c318c92f0c5→ HTTP 200; root 404.jsonspack.com→ NXDOMAIN. - Prior-cluster context on-wiki: algamil7x page third-sweep section (
chai-as-indexed/chai-as-viem, GHSA-c67f-wr24-prw9 critical, promotion bar); WeaselBiscuit (fetch→new Functiontwin).
Related pages
- Equation of Compromise campaign page, same-day discovery: the mathjs-cluster proves the audience-profiling half of this cluster's shape
- algamil7x npm DNS-exfil recon cluster, week's other install-script family
- Telegram chat-1064260758 dependency-confusion recon fleet, same-batch Inspector output, third rail (Telegram) of the same month
- WeaselBiscuit npm stealer, dead-drop fetch pattern ancestor