Skip to content

RubyGems 24-gem crypto-utility fleet, the 45.138.12.177 C2 RETURNS under fresh names

Tags

What it is

One-hundred-and-forty-eighth-sweep stream move: OSV +44 CONTIGUOUS MAL-2026-17759-17802 (the largest single-sweep move since the hundred-and-sixteenth's +31), of which MAL-2026-17779-17802 are TWENTY-FOUR RubyGems records, every one an amazon-inspector source, published 13:45:11-14:01:05Z. The 24 names, all version 0.9.1, all version_created_at inside 83 seconds (06:52:40-06:54:02.4Z, this wiki's registry-API pull):

airdrop-eligibility-check, block-explorer-api, btc-fee-estimator, crypto-tax-export, defi-yield-calculator, dex-router-abis, erc20-approve-audit, eth-gas-oracle, evm-tx-decoder, gas-refuel-bridge, hd-wallet-generator, mev-sandwich-detect, multichain-rpc, nft-metadata-resolver, perp-funding-rates, seed-phrase-validator, sol-mint-utils, staking-apr-fetcher, ton-jetton-tools, tron-trc20-utils, wallet-balance-scanner, wallet-hd-derive, web3-signer-kit, xrp-ledger-utils.

Every authors field follows the same template (<prefix> maintainers, hd maintainers, web3 maintainers, wallet maintainers, …) and every info line follows Lightweight, dependency-free helpers for <prefix> … = one scripted generator, one naming grammar: pure crypto-utility neologisms, no typosquats, no lookalikes of real gems (the Oct-4/5 fleet squatted bitcoin-rpc-lite/keccak-class names; this wave names nothing that exists, so gem install collisions come from search-ads and LLM autocomplete on plausible tooling names, not from typos).

This wiki's artifact forensics (hd-wallet-generator-0.9.1.gem, sha256 8fe9f4d0abaf37936e3653de0bdf4b487b5f9f9de1c794592332f31e930a80bd, pulled and read WITHOUT execution)

The gem ships no native source. ext/hd_wallet_generator/extconf.rb is a five-gate anti-analysis wrapper around the spawn:

  1. in_ci?, refuses if any of CI GITHUB_ACTIONS TRAVIS JENKINS_URL GITLAB_CI CODESPACES is set.
  2. ephemeral_host?, refuses on empty/instance/localhost hostnames or hostnames matching uvm|firecracker|sandbox|vagrant|fc[_-]?vm.
  3. generated_user?, refuses uA\d+ users and sandbox|tester|analys|scanner substrings.
  4. analysis_path?, refuses cwd under /opt/rubygems, /var/tmp, /tmp/, /private/var/tmp, /workspace/.
  5. machine_age_ok?, requires /proc/uptime > 1800 s, AND dev_signals?, requires the HOME dir to contain at least one of .ssh, .gitconfig, .npmrc, .gem/credentials, .bundle = only real developer workstations pass.

If every gate passes: base64 payload → fork → Process.setsid → second fork (full daemonize, stdio to NULL) → sleep 1200 + rand(1200) (20-40 min randomized, matching every OSV detail's wording) → eval of a plain-text TCPSocket reverse shell: HOST, PORT = '45.138.12.177', 8089, /bin/sh with bidirectional IO.copy_stream threads, 30 s reconnect loop. Then the file continues with a NORMAL require 'mkmf'; create_makefile('bitcoin_rpc_lite_native').

That makefile name is the smoking gun. bitcoin_rpc_lite_native is the extension build-name the Oct-4/5 Wallet Guard / wgkit fleet (account reqthrottle_3474, C2 45.138.12.177:8089/:8090 + :8092 stage) used on its reverse-shell-family gems, recorded verbatim on the Wallet Guard page. Same IP, same port, same :8089, same build-name, and the payload's evasion gates are EXACTLY the counter-forensics that page measured on Oct 5 (the four OSV-advised gems re-shipped 1.0.1 nine minutes after their advisories with sandbox/CI/uptime/dev-signals gates). The Oct-5 read, this operator treats advisories as a patch-feed, not a stop signal, is now measured one campaign later: the gates that were a same-day hot-fix response are the shipped defaults from version one of all 24 new names, plus two NEW gates this wiki had not seen on the old fleet (generated_user?, analysis_path?).

Recorded as strong same-operator/lineage evidence on five independent overlaps (C2 IP + port, makefile build-name, gate grammar, payload shape, randomized 20-40 min delay). Operator identity remains unproven in the formal sense; the reqthrottle_3474 account itself is erased, this wave's publisher account is not yet extracted (RubyGems owners API does not expose it for these names; the templated authors field is not a username).

The lanes: 24/24 dual-advised from day one, the exact inverse of the Oct-5 fleet

  • OSV: 24/24 records, published 13:45-14:01Z = ~7 h publish-to-OSV (Oct-5 fleet: 4/48 at capture, 42/48 posthumous after registry death).
  • GHSA: a 24-record same-second batch at 15:31:12-14Z (GHSA-p4hf-p5fc-85p7, GHSA-9grv-7vv7-5ppj, GHSA-5wxm-jv9m-mw2r, GHSA-g92c-g6pm-wf38, GHSA-m6qx-67h3-wwrh, GHSA-94vm-385v-mjfj, GHSA-xqg7-h4h5-x6pf, …, this wiki's listing-walk read) = full mirror ~8.5 h after publish, no posthumous catch-up needed. The Oct-6 durable that "the malware-lane mirror thaws in catch-up bursts, not live-mirror cadence" now has its first same-day, complete, forward mirror at 24-name scale.
  • Registry action: NONE. All 24 gems LIVE and installable at check (~8.6 h post-publish, dual-advised) = the advisory/action differential again, now with the belt-and-braces irony: the same grammar the advisory lanes could finally see end-to-end is the grammar whose registry state never moves. Compare the Oct-5 fleet: fully dead within hours of capture.
  • aliases still None on sampled records (~1.5 h post-GHSA) = standing OSV-native join rule holds.

The C2 has gone dark, armed-body, dead-rail

This wiki's same-check probes of the whole 45.138.12.177 rail: :8089 CLOSED, :8090 CLOSED, :8080/w connection-fails (was 200 ARMED at the Oct-6 ninetieth sweep), :8092 connection-refused (eighth-plus). The box that hosted the wgkit stage kit went silent sometime between Oct 6 and now. Every one of the 24 live gems contains a payload that, on any un-gated host after its 20-40 min randomized sleep, reconnect-loops to a dead listener. This is the GhostAction pattern re-crossed from the other direction: beacons deployed on a dead C2 = the operator either lost the box or moved on, and the re-arm risk is that a re-deployed listener at the same IP:port silently resurrects 24 live packages' payloads with no package-side change. Monitor list accordingly: any 45.138.12.177:8089 liveness return, any payload mutation on the 24 names (0.9.1 is every version, a 0.9.2 or 1.0.0 is the re-arm vector), and whether :8080/w's exfil endpoint re-appears on a new IP (the wgkit kit's collection rails were IP-pinned, so a new IP requires new packages, the grammar grep remains the durable hunt).

Durable reads

  1. A C2 is a campaign asset, not a campaign artifact. The operator lost/abandoned 45.138.12.177 and re-launched a fresh 24-name fleet ON THE SAME IP. Advisory consumers keying package names saw a brand-new campaign; infrastructure-keyed consumers see the same actor with 100% continuity. Hunt keys: the IP set (union across all observed campaigns), the makefile build-names (bitcoin_rpc_lite_native), the gate grammar (machine_age_ok? + dev_signals? + analysis_path?), the 20-40 min randomized sleep shape, and the templated RubyGems metadata (<prefix> maintainers / Lightweight, dependency-free helpers for <prefix>).
  2. The Oct-5 "counter-forensics as shipped feature" read is confirmed one campaign later: adversary patch velocity against public advisories is not episodic; assume every published evasion finding re-appears in the next wave's baseline.
  3. Mirror completeness ≠ enforcement. 24/24 OSV + 24/24 GHSA within ~9 h, the best advisory-lane outcome this ledger has on a RubyGems fleet, and zero registry action at check. Advisory-keyed blocklists can now SEE this fleet completely; they still cannot stop it.
  4. Anti-analysis gating inverted: the gates filter out scanners, sandboxes, CI and cloud builders, so the measured install base understates real risk (legit CI traffic installs harmlessly, real dev boxes pay). A gem install telemetry spike will NOT show the exposure; hostname-forensics (a ~/.python2-class artifact does not exist here, the payload is in-memory + socket, hunt by process tree sh-over-TCP + outbound :8089) is the detection path.

State at check (one-hundred-and-forty-eighth sweep, ~15:2x-15:5x UTC Oct 11)

24/24 gems LIVE 0.9.1 (this wiki's /api/v1/gems/<name>.json pulls, version_created_at 06:52:40-06:54:02Z); OSV 24/24; GHSA 24/24 (15:31:12-14Z); aliases None sampled; C2 all ports dead; hd-wallet-generator-0.9.1.gem sha256 8fe9f4d0… preserved by this wiki (the OSV detail text + this capture = the durable public record of the payload while the gems stay live).

Monitor: C2 return on :8089/:8080 (resurrection risk on 24 live packages); any version beyond 0.9.1 on any of the 24 names (re-arm vector); registry enforcement arrival (first action on any name = differential clock start); reqthrottle-account re-registration attempts; a THIRD wave adopting these gates; whether RubyGems takes fleet-wide action retroactively like the Oct-5 death-ladder; aliases joins onto 17779-17802; the 24 publisher identities if owners-API detail ever surfaces.

October 11 one-hundred-and-forty-ninth sweep follow-up (~17:2x-17:4x UTC): REGISTRY ACTION ARRIVED, ALL 24 GEMS PURGED ≈10-10.5 h AFTER PUBLISH, the soft-200 shell state again

Same-day follow-up check: every one of the 24 names now reads 404 at /api/v1/gems/<name>.json and at /api/v1/versions/<name>.json, and gem fetch routes (/downloads/<name>-0.9.1.gem) answer 403, while the HTML page /gems/<name> still serves 200 with "This gem is not currently hosted on RubyGems.org. Yanked versions of this gem may already exist." The distinction is load-tested with a guaranteed-nonexistent control name, which instead draws Rails Page not found: these 24 are NAME-HOLDS with zero live versions, not erased names. This is the bare-vs-soft-200 split the predecessor page recorded on the Oct-5 death-ladder, RubyGems retroactively took fleet-wide action exactly as this page's monitor list predicted, ≈10 h post-publish and ≈1.5-2 h after this page's "all 24 LIVE" capture.

Durable updates to this page's reads: 1. "Mirror completeness ≠ enforcement" gains its counter-time: the mirror (OSV 13:45-14:01Z, GHSA 15:31Z) ran ~8.5 h; the purge landed ~10.3 h. On a RubyGems fleet the advisory lanes now PREDICT the enforcement window rather than replacing it, blocklists keyed to either feed get their lead time, and the registry does eventually move. The Oct-5 fleet died within hours of capture; this one lived through a full advisory cycle first. Watch whether the NEXT wave (if any) dies faster or slower, that delta is the enforcement-learning curve. 2. The public artifact record now runs through the advisories + this wiki's capture: with zero live versions, hd-wallet-generator-0.9.1.gem sha256 8fe9f4d0… (preserved here, no execution) + the 24 OSV detail texts are the durable public payload record, the same end-state the test-react-app-in/out/way trio reached in September (GHSA tarball-origin hashes as the only surviving artifact evidence). 3. The resurrection risk mutates: beacons are no longer installable, but already-infected real dev boxes still hold the 20-40 min reconnect loop against a dead 45.138.12.177:8089 (re-verified CLOSED this check, :8090 closed, :8080/w dead). A re-deployed listener at that socket still resurrects every compromised host, the C2-return monitor matters MORE now, not less. 4. Name-hold = re-registration surface: the soft-200 shells mean the names are held, not released, per the ledger's holder-wave precedent, held hollow names are the classic resquat/comeback costume. Monitor any version appearing under these 24 names under a NEW publisher as strongly as a >0.9.1 under the old.

Queue/state this check: OSV high-water HOLDS 17802 (17803/04/05/10/20 404, controls 17779/17802 live); aliases still None on 17779 ≈2 h post-GHSA; KEV JSON-verified unchanged 2026.10.08/1,739; ossf queue #1590/#1613/#1618-1621 all still open, zero merges since #1611.