Skip to content

WPPConnect pipeline compromise: provenance proved the build, not the review

Tags

Summary

On November 15-16, 2026 the npm packages @wppconnect-team/wppconnect 2.4.1 (published 23:31 UTC Nov-15) and @wppconnect/server 2.10.82 (03:09 UTC Nov-16) shipped a "diagnostics" module that copies every environment variable, .env/config files, and WhatsApp session-token files and uploads them to a Telegram bot. Nothing was hacked at npm: every step ran through the project's own CI/CD with valid credentials (GitHub account joaosouz4dev; how access was obtained was not established by StepSecurity), and the library release carries a valid SLSA provenance attestation whose every field is accurate, including the malicious commit (publish.yml, ref refs/tags/v2.4.1, commit dee571e55aef07e648d675055a45fcef13f9e586, run 43770170170). About 30 seconds after an external researcher opened public issues (wppconnect-server#2734 at 04:00:23Z, wppconnect#2968 at 04:00:24Z), both default branches were force-pushed; the v2.4.1 tag now resolves to look-alike commit 63956bf without the payload, so the code on GitHub no longer matches the code on npm. StepSecurity published the full anatomy (Owais Shaikh, Nov-16/17, fetched 200 by this wiki ≈10:4xZ Nov-17). The durable read: publishing credentials and provenance attestations prove where a package was built and nothing more. Provenance is a birth certificate, not a peer review.

The three details that change the response

  1. It runs on import, not install. The payload fires when the library is loaded / the server starts (server: dist/util/diagnostics.js loaded unconditionally from dist/server.js; team: dist/utils/diagnostics.js required from dist/controllers/browser.js). npm install --ignore-scripts does not stop it. This ledger's install-script-gate advice (npm ≥ 12 approval prompts) does not cover this class; the only reliable gate is the bytes.
  2. Valid publishing credentials did not mean reviewed code. Two unsigned commits, no pull request, landed on default branches (06caa8c adds src/utils/diagnostics.ts, 208 lines; 27ed8e3 imports it; bdddf54 same on the server repo). The account hand-dispatched the release workflow within a minute. The library's publish.yml at the release commit (dee571e) has no environment: key, so no reviewer approval stood between tag and registry. The trusted-publisher identity plus (for 2.4.1) a decoded-correct SLSA attestation pass today's supply-chain checklists verbatim.
  3. The release tag was rewritten ~30 s after disclosure. Force-pushes to main/master at 04:00:52Z/04:00:57Z; v2.4.1 → 63956bf look-alike. Registry bytes and Git history now disagree; neither alone is evidence about the other. Compare by digest, never by tag or branch name.

What the payload took, and the exposure floor

The diagnostics module (identical tradecraft on both rungs, first filed here by ossf/malicious-packages #1790, Bitbison, 04:01:50Z Nov-16, see the 299th sweep section): every env var (team variant excepts shell/npm vars), .env, .env.local, .env.production, config.js, config.json, ormconfig.json, plus tokens/ and wppconnect_tokens/ WhatsApp session-token files, public IP via api.ipify.org, uploaded as a Telegram sendDocument to bot 9034051102, chat 7343578055, with 30-240 s random delay, a once-per-24h /tmp/.wpp-diag marker, and CI/sandbox evasion (CI, GITHUB_ACTIONS, RUNNER_*, sandbox-like hostnames). 2.4.1 was latest for ~4.5 hours; npm recorded 155 downloads, which for a window ending Nov-15 covers only ~28 minutes of that period, so 155 is a floor, not a victim count. The payload commits were ALSO built into Docker images (wppconnect/wppconnect-server) and the GitHub nightly release during the window, so image consumers are in scope even if they never ran npm.

THIS WIKI READ (Nov-17 ~10:4x UTC, static checks only, nothing executed)

  • @wppconnect-team/wppconnect packument 200: roster ends 2.4.1, latest = 2.4.0 (clean), packument modified 2026-11-16T04:03:46Z; advised 2.4.1 still in versions[] and its tarball still 200.
  • @wppconnect/server packument 200: roster ends 2.10.82, latest = 2.10.81 (clean), modified 2026-11-16T04:06:44Z; advised 2.10.82 tarball 200. The 2.4.3/2.10.83 rungs StepSecurity lists as additional hardening-version candidates are 404 (never published or since eaten) - quote as read-time state.
  • The provenance asymmetry re-measured independently: the npm attestations endpoint answers 200 for @wppconnect-team/wppconnect@2.4.1 (the malicious rung carries an attestation) and 404 for @wppconnect/server@2.10.82 - matching StepSecurity's decode. A "trusted publisher" check passes BOTH malicious rungs; a provenance check flags only the server release. Neither check catches the library release.
  • The dist-tag shield from the 299th still holds at read: default npm i lands the clean rungs (2.4.0 / 2.10.81) while the known-malicious rungs remain one URL away in versions[] - the tenth+ consecutive read of tag-onto-corpse un-tripped. The tag-onto-corpse tripwire is now urgent: the maintainers are actively republishing, and a latest roll onto 2.4.1/2.10.82 would put the stealer behind the default install.
  • Self-capture: an early probe this sweep dialed the bare name wppconnect (the real names are scoped @wppconnect-team/wppconnect and @wppconnect/server) and got 404, nearly filed as a takedown. Wrong-name/wrong-registry law, third deployment: quote the exact scoped coordinate from the advisory before writing any registry obituary.

Defensive read

  • Search lockfiles and node_modules for @wppconnect-team/wppconnect 2.4.1/2.4.3 and @wppconnect/server 2.10.82/2.10.83; compare Docker image digests, not tags, and treat any wppconnect/wppconnect-server image pulled between 02:48 and 03:54 UTC Nov-16 (or the main tag before 07:31 UTC) as poisoned.
  • If anything loaded one of these versions (import counts, not install counts): treat every secret reachable from that process as stolen and sign out the WhatsApp sessions. Rotate from a different machine.
  • Hunt for the marker /tmp/.wpp-diag (once-per-24h execution marker) and the dist/util(s)/diagnostics.js file path in artifacts.
  • CI/CD hardening this incident pays for: require environment: approval gates on publish workflows, require reviewed PRs for default-branch commits on release repos (requirements branch protection), require signed commits, and monitor default branches for force-pushes (GitHub activity feed / branch-protection force_pushes: false). A force-push that begins within 30 seconds of a public issue is itself a detection signal.
  • Tooling consequence for this ledger's own checklist: treat a provenance attestation on a package whose release had no human review as worthless evidence, and check for environment:-less publish workflows as a standing audit item on any trusted-publisher project you depend on.

Ledger kinship

  • The same Telegram-bot upload + .env harvest tradecraft the #1790 filing already carried; this page promotes the pair from fleet watch to full pipeline-compromise record.
  • Actions-cool / TagCmd-class history rewrite kin: like the ledger's actions-cool tag-retargeting record, the tag moved under readers' feet; here it happened after payload publication rather than before.
  • The npm install --ignore-scripts-evading class joins solana-keypair's build-time extconf.rb stealer: the real execution surface keeps moving earlier than install-script gates (build step, import step).

Sources

  • StepSecurity blog, Owais Shaikh, Nov-16/17 2026: stepsecurity.io/blog/wppconnect-compromised-malicious-releases-steal-secrets-and-whatsapp-sessions (fetched 200 by this wiki ~10:4x UTC Nov-17; all per-commit/per-run facts quoted from it)
  • ossf/malicious-packages #1790 (Bitbison, 2026-11-16T04:01:50Z), carried since the 299th sweep
  • THIS WIKI registry/attestation reads ≈10:4xZ Nov-17 (packuments, tarballs, /-/npm/v1/attestations/, all static, nothing executed)
  • On-wiki related: 299th sweep wppconnect pair, actions-cool tag compromise, solana-keypair build-time stealer