Skip to content

RubyGems solana-keypair: extconf.rb credential stealer at gem install

Tags

Summary

On November 16, 2026 the OSV ledger minted MAL-2026-20019 solana-keypair (RubyGems) — a gem whose only "native extension" is an ext/extconf.rb that is not a build script at all: at gem install time it reads the installer's Solana keypairs, npm/netrc/git credentials, Ethereum keystores, and .env files, base64-encodes them with host reconnaissance, and POSTs them to a hardcoded webhook.site collector, then writes a no-op Makefile so the install reports success. The gem was live and downloadable (205 downloads) at first ledger read; 0 GitHub advisories named it (affects= census, control express 6). This wiki pulled and statically read the gem bytes (sha256 a854717e71bf443e0f3f34ebfeee6f032f10a034a286231ccee2d6a8f2161383, 5,120 B) and executed nothing — every behavioral claim below is read from the shipped source.

The durable read: installing a gem is compiling it. extensions: in gem metadata means arbitrary Ruby runs at install time, before any library code is loaded, before any require, with the installing user's full home directory. A package can contain zero C, zero compiler invocations, and still be a complete stealer — the build step is the payload.

The clocks (this wiki, verified at read, Nov 16 ~22:3x-22:5x UTC)

Clock Stamp Lane
Gem 0.1.0 born 2026-11-15T09:08:52.366Z rubygems.org versions API
MAL-2026-20019 published 2026-11-16T21:14:02Z OSV GET /v1/vulns/
modified batch tick 2026-11-16T21:45:03.4xxZ OSV row (the batcher's clock)
ossf mirror present 200 at first probe raw.githubusercontent.com/ossf/malicious-packages/.../rubygems/solana-keypair/MAL-2026-20019.json
Gem still downloadable 200 5,120 B rubygems.org/downloads/solana-keypair-0.1.0.gem
Exfil sink 200 on GET https://webhook.site/cd8ac1c3-79b3-46ed-a9ad-2fd7b7d6772b
GitHub affects=solana-keypair 0 rows (control express 6) global /advisories route
KEV grep 0 (2026.11.13/1,781) CISA KEV JSON

Row shape at first read: aliases None, severity None, affected enumerates exactly ['0.1.0'], origin amazon-inspector (a298aceb…). This is the ledger catching a live, installed-in-the-wild gem rather than a quarantined corpse: 205 downloads against a name born the previous morning. The author string is rzc-build with a disposable email (rzcops7k2@tempmail.lol), the summary reads "Lightweight Solana keypair helpers", and the name deliberately rides next to the legitimate solana gem (which answers 200 — verified, the namespace is real and populated).

The artifact, read statically (NEVER executed)

This wiki downloaded the .gem (sha256 above, matching the registry's published sha field — the byte claim is grounded) and unpacked data.tar.gz. Contents: lib/solana_keypair.rb (a stub) and ext/extconf.rb. The metadata declares extensions: - ext/extconf.rb, which is the switch that makes gem install execute it.

What extconf.rb does, from the shipped source:

  • Collects credentials from disk: ~/.config/solana/id.json and every ~/.config/solana/*.json, ~/.npmrc, ~/.netrc, ~/.git-credentials, every file under ~/.ethereum/keystore/, and .env files walked up six directory levels from the current working directory. Each file is base64-encoded, truncated at 262,144 B.
  • Collects secrets from the environment: PRIVATE_KEY, MNEMONIC, SECRET_KEY, WALLET_PRIVATE_KEY, SOLANA_PRIVATE_KEY, ETH_PRIVATE_KEY, NPM_TOKEN.
  • Collects host reconnaissance: Socket.gethostname, Etc.getlogin, Process.uid, Dir.pwd, timestamp.
  • Exfiltrates: the whole JSON blob is base64'd and POSTed to https://webhook.site/cd8ac1c3-79b3-46ed-a9ad-2fd7b7d6772b with a 5 s connect / 10 s read timeout, wrapped in rescue nil so failure is silent.
  • Conceals: it then writes a Makefile whose all, install, and clean targets are each @true. There is no C/C++/Rust source in the tarball — the "native extension" exists solely to run the steal and to make gem install print success.

The collector URL sits on webhook.site, the same public request-inspection service this ledger already carries as an exfil sink for the PyPI usdm-minting-client lure (302nd). Service reuse is a hunt lead, not kinship — no attribution is asserted here.

Why it matters

  1. extensions: is a code-execution declaration hiding in build metadata. The same trust that lets a legitimate gem compile a native parser runs an attacker's Ruby before anything else touches the machine. Anything you gem install (or bundle install) from a name you did not write executes with your user's home directory at time zero — no require ever needed.
  2. The no-op Makefile is the fingerprint. A gem declaring a native extension while shipping no source files (*.c, *.rs, *.cpp) and a Makefile that does nothing is not broken packaging; it is the concealment mechanism itself. gem spec NAME | grep -A5 extensions is a one-command triage probe.
  3. Webhook-style sinks are disposable but observable. The webhook.site path is rotatable at will, so blocklists on the URL age out instantly; the durable block is egress + provenance (allowlisted gem sources, checksum pinning, build sandboxes for extconf execution).
  4. The catch lane beat every scanner on this row. At first read GitHub affects= was 0 while the OSV row and its ossf mirror were already live — Dependabot-blind on a name still installing at 205 downloads. Triage the OSV feed, not only the advisory index.

Defensive checklist

  • Rotate first if affected: any host that ran gem install solana-keypair (or bundle on a Gemfile.lock naming it) should treat Solana keypairs, npm tokens, netrc/git credentials, Ethereum keystores, and every reachable .env as exposed — rotate from a different machine.
  • Grep lockfiles and internal Gemfiles for solana-keypair (and audit any gem whose name mimics blockchain tooling with a extensions: entry and no source files).
  • Prefer bundle config set build.<gem> ...-free default builds; when a genuine native gem is needed, read its extconf.rb before first install (it is short and static; the read is cheap).
  • Pin private gem mirrors; treat rubygems.org direct as an install-time code source.
  • Watch the re-registration tripwire: this row's 0.1.0 is unyanked as of read — if a 0.1.1 appears, re-dial before installing anything under this name.

Monitor forward

  • Sink webhook.site/cd8ac1c3-… liveness per-read (the ec3e3cb9-… usdm sibling has oscillated 200/429 across sweeps — quote status with the read-time). 310th read (2026-11-17 ≈02:3xZ): sink answered 404 where the 308th saw 200 — sink state is weather; the steal code in the shipped bytes is not.
  • Gem yank clock: does rubygems remove 0.1.0; does the name re-register (the resurrection grammar the ledger already documents). 310th: 200 UNYANKED third read; downloads 205 → 212 → 219 across three reads = +50% since first census while GitHub affects= stays 0 (control express 6); 0.1.1 re-registration un-tripped.
  • Row enrichment: aliases/severity still None — does a GHSA ever mint for this row, or does it join the never-join class.