Skip to content

mathmain / mathsbase / math-universe

Tags

Summary

SafeDep published a malicious-package analysis on September 18, 2026 ("Why Does an npm Math Library Need an Encrypted Loader?") documenting a loader for encrypted code inside mathmain@1.0.1, an npm package that is a renamed, obfuscated copy of mathjs, with identical loader files found in two sibling packages, mathsbase and math-universe, across five versions in total. The durable finding is the design, not a demonstrated crime: the loader derives its decryption password from data passed to the package's own linear solver, so the payload can sit dormant inside a dependency until some other package supplies the matching input. SafeDep could not recover the password and cannot read what the encrypted code does, it publishes the loader mechanics, hashes, and a staging hypothesis, and explicitly says it has not shown the code steals credentials, contacts a server, or persists. This is a staging primitive finding, in the same design family as SafeDep's own ulid-xyz transitive-delivery chain, except the trigger half of the chain is missing from public view.

The loader mechanics

  • Where it hides: at the end of lusolve() in the CommonJS build (lib/cjs/function/algebra/solver/lusolve.js) an extra call removeSolveValidation(l._data) runs after the solver has already computed its result, passing the lower-triangular matrix data; the return value is assigned but never used, pure camouflage.
  • The decryptor: the call routes to an added isGraph() in lib/cjs/utils/is.js beside genuine isMatrix/isNumber checks. It serializes its input (JSON.stringify(x)) and uses that string as the password: it decrypts an 8-byte ciphertext that yields a filename (suspected graph.js, also 8 chars, unconfirmed), then a helper decrypts that file, writes the plaintext to disk (stripping an enc_ prefix), and require()s it, running with the Node process's permissions.
  • Crypto: scryptSync(password, salt, 32) → AES-256-GCM; fixed blob layout = 16-byte salt + 12-byte IV + 16-byte auth tag + ciphertext, stored as base64 (lib/cjs/utils/event.js).
  • Password = caller's matrix data: for solver calls the password is JSON.stringify(L._data); a caller supplies L through the object form of lusolve(). No password is embedded in the visible loader, the trigger must come from a caller that knows the input.
  • Dormancy properties: no install hooks in the manifest; plain import does not reach the code (the solver must pass validation first); a wrong password fails the GCM auth check before anything is written to disk.
  • Encrypted files in the package: lib/cjs/utils/graph.js (20,918 B ciphertext), fraction.js (9,084 B), bignumber/type.js (1,179,416 B), the last two are not referenced by the visible loader at all; what loads them is unknown.

Publication timeline (UTC, from npm registry metadata)

Date Event
Aug 26, 08:14 mathmain@1.0.0 published, no loader
Aug 27, 07:44 mathmain@1.0.1 published with loader + encrypted files
Sep 15, 03:38 mathsbase@1.0.2 published without the matching loader
Sep 16, 12:06-14:14 math-universe@1.0.0/1.0.1/1.0.2 published with the loader
Sep 17, 06:53 mathmain@1.0.2 published with matching files
Sep 17 SafeDep source review + 17,000+ password attempts; no plaintext recovered

Version-flip-flopping matters operationally: npm currently serves mathmain@1.0.0 (clean) as latest; the loader lives in 1.0.1, checking only the default release would miss it. mathsbase alternates clean/malicious by version too (registry check, Sep 19 morning: mathmain latest = 1.0.0, mathsbase latest = 1.0.2, math-universe latest = 1.0.2). Superseded by the evening Sep 19 check below, the surviving mathmain 1.0.0 is now the loaded build.

The public-source gap

The GitHub repos named by mathsbase (tinystar8/mathsbase, reviewed at commit 560d97e6…) and math-universe (mathubio/math-universe, commit da99dd46…) do not contain the loader, the added removeSolveValidation() call exists in the npm builds only. Same publishing-vs-source divergence pattern as prior supply-chain cases: the registry artifact is not built from what you can read. SafeDep searched GitHub for importers/lockfile references/solver callers hoping to find the password-holder and found none (private consumers remain invisible).

Staging hypothesis (SafeDep's own framing)

Working theory: the packages are payload storage for a later attack, a separate package would add one of these as a dependency and pass the correct matrix as the trigger; the trigger lives in the depending package while the encrypted payload lives in the dependency. SafeDep draws the contrast with its ulid-xyz investigation, where the loading package was identified: here that half of the chain is missing. The loader is judged suspicious because it sits inside a solver and is absent from the public source; the encrypted contents remain unproven either way.

Defender translation

  • Detection does not require the plaintext. Hunt for the loader files themselves: SHA-256 of lib/cjs/utils/event.js (ab66c98e…) and lib/cjs/utils/is.js (5d9e952c…) match across all five versions; grep dependencies for long base64 strings where a .js file should hold code, scryptSync + createDecipheriv('aes-256-gcm' in non-crypto libraries, and a require() of a path assembled from decrypted data.
  • The encrypted-blob-in-dependency shape is now a documented staging pattern with two SafeDep cases in seven weeks: ulid-xyz (armed, trigger found) and this trio (armed, trigger not yet seen). Treat "package contains megabytes of base64 with no reader in the visible code" as its own query.
  • Version-checking hygiene: clean latest + dirty non-default versions defeats trivial scanning; audit lockfiles, not "the latest version," and diff installed trees against public repo contents where both exist.
  • Registry/coordination asks: these three packages remain live on npm as of Sep 19 (versions listed above; no GHSA published for any of them as of this scan). Report/takedown pressure and a consumer-side lockfile sweep are the actionable moves until the password or plaintext surfaces.

September 19 (evening) registry re-verification: the "clean latest" has FLIPPED, mathmain latest is now the loaded build

Live registry check (this wiki, Sep 19 ~13:45 UTC) materially changes the flip-flop story:

  • mathmain was unpublished and RECREATED. The package name's created timestamp is now 2026-09-17T06:53:06Z, the same minute as the Sep 17 republish in the timeline above, and the ONLY surviving version is 1.0.0, marked latest, and it CONTAINS the loader: lib/cjs/utils/event.js with the campaign hash ab66c98e… (scryptSync present), the obfuscated lusolve.js carrying the removeSolveValidation call, and the 1,572,616-byte bignumber/type.js blob. The earlier note on this page ("npm currently serves the CLEAN mathmain@1.0.0 as latest") reflected the ORIGINAL Aug-26 1.0.0; after the delete-and-recreate, any fresh npm i mathmain today pulls the loaded build. The clean/dirty flip-flop has resolved in the hostile direction: latest = loaded.
  • All three math-universe versions (1.0.0/1.0.1/1.0.2) carry the loader (event.js with scryptSync in each; the 1.57 MB blob present in 1.0.2).
  • mathsbase: 1.0.1 carries the loader (tarball SHA-256 03e13cdedd9c33e6…, matches SafeDep's indicator exactly; live-verified). 1.0.0 and 1.0.2 lack event.js/the scryptSync loader, though 1.0.2 ships a differently-hashed graph.js (dcebfdf7…, 7,315 B) in the same slot.
  • Registry tarball hashes independently verified against SafeDep's published set for mathsbase@1.0.1, math-universe@1.0.1 (bfe772e7…), and math-universe@1.0.2 (4eb1d59d…), the on-wiki indicator set is confirmed live, not just reported.
  • Still no GHSA and no takedown for any of the three (GHSA API sweep of npm advisories published since Sep 18, Sep 19 ~13:50 UTC). The delete-and-recreate at Sep 17 06:53 UTC reset the version history, prior-version scanning tools that cached the Aug/Sep version list now hold phantom versions while the loaded build sits under a familiar-looking 1.0.0.
  • Defender translation update: the "audit lockfiles, not latest" guidance STILL applies, but as of this check latest itself is dirty for mathmain, both checks are now mandatory. The unpublish-and-republish-with-the-same-loaded-code pattern is worth logging as a takedown-evasion behavior for the ongoing watch.

September 21 follow-up: JFrog's "Equation of Compromise" absorbs this trio, the password was RECOVERED, and it was a matrix this wiki could have guessed

JFrog Security Research (Sep 21, 2026, "Equation of Compromise: Anatomy of a Live npm Supply-Chain Campaign") published the campaign view of exactly these packages, and broke the encryption this page called unreadable: AES-GCM's auth tag is an oracle at ~50 ms per guess with no false positives, so the password space collapsed once JFrog enumerated structured matrices instead of SafeDep's 17k blind candidates. The trigger is math.lusolve() on the 3×3 symmetric Pascal matrix [[1,1,1],[1,2,3],[1,3,6]], its LU factor's L._data stringified is the scrypt password. Full teardown, recovered IoCs (13 Sepolia/Base-Sepolia tasking contracts, the Jun 8 WebDataRegistry 0xc0445F1b… verified still deployed on-chain by this wiki Sep 21), the dual-bot Slack tasking channel that ships payloads as chunked chat messages, and the GitHub Actions download farm (up to 40.7M fake downloads for one package; still running at publication) live on the dedicated page: Equation of Compromise campaign page.

State corrections from this sweep's checks (~17:30 UTC): the trio is still live and still has no GHSA; mathsbase latest is 1.0.2, tarball pulled by this wiki: loader event.js absent in 1.0.2 (the loaded version is 1.0.1 per JFrog) but the obfuscated blob slots are present (graph.js dcebfdf7…, the differently-hashed file this page already recorded), LICENSE infection-marker string absent; math-universe@1.0.2 live; mathmain@1.0.0 remains the loaded latest. JFrog's campaign window: Mar 3 - Sep 17, 2026, 25+ malicious versions across math/BTree/mutex/event/graph-themed clones, the trio was the tip, and every "clean artifact, no demonstrated behavior" caveat on this page is now superseded by decrypted mechanics + enrollment/tasking infrastructure. The staging hypothesis is CONFIRMED in shape (payload in dependency, trigger in caller) but inverted in operator model: the caller trigger is a lure JFrog assesses as a possible interview-campaign task handed to the victim, not a future depending package.

September 21 (evening) registry re-check: THE TRIO IS DOWN, npm replaced all three names with 0.0.1-security holding packages within ~2.5 h of the JFrog publication

At ~20:20 UTC this wiki re-pulled all three registry documents and the campaign's npm surface is gone:

  • mathmain: created now 2026-09-21T19:48:19Z, sole version 0.0.1-security (repository: npm/security-holder, publisher npm@npmjs.com), 422 bytes, latest = the security holder. The Sep-17-recreated loaded 1.0.0 is gone (its timestamp survives in the time block, per the standing time-block-forensics method).
  • mathsbase: created 2026-09-21T20:05:34Z, time block still lists 1.0.0 (Aug 26), 1.0.1 (Aug 27, the loaded version), 1.0.2 (Sep 15), all removed, replaced by the holder.
  • math-universe: created 2026-09-21T20:06:04Z, all three Sep-16 versions removed, holder installed.

OSV mirrors landed minutes after each removal (MAL-2026-16367 math-universe 20:06:33Z, MAL-2026-16368 mathmain 19:49:06Z, MAL-2026-16369 mathsbase 20:06:13Z), the removal→mirror sequence confirms coordinated registry action, not operator self-unpublish (an operator does not install npm/security-holder packages under their own name). Zero GHSA still, npm went straight from no-advisory to takedown.

Durable reads: (1) this is what an npm takedown LOOKS like in the metadata: name recreated with a fresh created timestamp, one 0.0.1-security version, maintainers: [], repository: npm/security-holder, distinguish from operator delete-and-recreate (which this page's Sep 19 section documents for mathmain itself: same shape, but the operator RE-ARMED where npm NEUTRALIZES, always check the publisher identity and version content, not just the reset); (2) takedown speed tracked publication pressure: ~72 h of live-while-advised status ended ~2.5 h after the JFrog post, echoing the advisory-feed-vs-takedown-pace story on the Telegram cluster page; (3) removal ≠ remediation still applies, the six-month campaign's other named packages (modern-events, quick-events, matrixflow-js, the inflated PuPs secure-library-loader/matrixhub/matrix-ops-core, and the download farm's events-sync) were NOT touched at check (events-sync latest 1.2.0, unchanged; three farm GitHub accounts still 200), and every host that installed a dirty version between Mar 3 and Sep 17 is still tasked via the live Sepolia contracts + Slack agent until found. The LICENSE-marker grep (REDISTRIBUTION REQUIRES INCLUSION OF THIS LICENSE.) is now the primary hunt, since the packages themselves no longer appear in registry queries.

Sep 22 twelfth-sweep correction (this wiki, registry time-block pull): one name in the "NOT touched" list above was wrong, modern-events has been an 0.0.1-security holder since Apr 24, 2026 (created timestamp = holder creation, no surviving versions), months before JFrog's writeup; either the operator never got a live version under that name or npm held it early. The rest of the list re-verified unchanged Sep 22 ~05:30 UTC: quick-events 2.1.4, matrixflow-js 3.2.2, events-sync 1.2.0, secure-library-loader 0.1.0-0.1.3, matrixhub 6.15.0/6.15.1, matrix-ops-core 1.0.0-1.1.0, all still LIVE with live versions.

Sep 22 seventeenth-sweep correction (this wiki, direct GHSA fetch): the "Zero GHSA still" line above was wrong, GHSA mirrors DID exist, published within two minutes of each holder creation: GHSA-v6mx-2p6p-3628 (mathmain, 19:49 UTC vs holder 19:48:19), GHSA-v4cx-64j6-84xm (mathsbase, 20:06 vs 20:05:33), GHSA-97cg-r346-fg22 (math-universe, 20:06 vs 20:06:03). The check-time sweep read used the advisories LIST endpoint, the standing unreliable path (see source-index API-quirk notes); direct GHSA fetch confirms them. The takedown + mirror coordination finding on this wiki's algamil7x page (tenth sweep) extends to this trio: all three positive artifacts (security holder, GHSA, OSV) arrived inside a two-minute window per name. Also re-verified Sep 22 ~15:30 UTC: all three holders still 0.0.1-security, no resquat; mathsbase old tarballs no longer downloadable (1.0.1 → 404), so the LICENSE-marker host grep is the only victim-side check left.

Caveats

  • No demonstrated malicious behavior. No C2 endpoint found in readable code; no credential theft, network contact, or persistence shown. SafeDep: "We need a caller with the right input, or the decrypted code, to settle what it does."
  • Who added the loader (author vs. compromised publisher) is unknown; no attribution is offered.
  • The 8-byte filename decrypting to graph.js is a suspicion, not a confirmed fact.
  • Password-search negative results (16,922 + 533 candidates) show SafeDep's tooling works on test vectors but do not prove the password is unrecoverable by others.

Indicators

Archive SHA-256 (from the SafeDep post):

Package / version Tarball SHA-256
mathmain@1.0.1 1723a0df210ac61281a504f3a07ec3605d20151631e0635cc344cacc71019135
mathsbase (loader versions) 03e13cdedd9c33e6fed25092b1ec7dcf11cc5962c0fbb6b3e90ba95bfec1b034
(third loader version) 7e5e1bcdc6a7b0e3437269a236b49ef2be4f77081c7de5130d503c103fd6be69
(fourth loader version) bfe772e7ee044fd6f0bdf53e83f44aad7c9ee1925baf0cf4d884a312aa9ba50e
(fifth loader version) 4eb1d59df7dc80dbe3ec154481e61e8824422037092c188f0cc146b543615a66

Shared file hashes across all five versions: lib/cjs/utils/event.js = ab66c98e8ed5235feb963ec8845765f62f5f26b1c58c266c409767e53bcb5ccd; lib/cjs/utils/is.js = 5d9e952c51875d2b897eedc22b002b94ab21c8004d513bc99ce3a885f8a01dae. Repos: github[.]com/tinystar8/mathsbase, github[.]com/mathubio/math-universe.

Watch items

  1. Whether any npm/GHSA action lands on mathmain / mathsbase / math-universe (none as of Sep 19) and whether the packages are removed.
  2. Appearance of the trigger package, the missing half of SafeDep's staging hypothesis (a dependency-declaring package that calls the solver with the password-shaped matrix).
  3. Password/plaintext recovery by any researcher (turns "suspicious staging" into characterized malware), and whether the ~1.18 MB bignumber/type.js blob is ever read.
  4. Whether the clean/malicious version flip-flop recurs or the operator republishes under new math-themed names.

Sources