Skip to content

orderly-plugin-liquidity-heatmap, a DEX-plugin costume whose postinstall profiles GitHub Actions…

TL;DR: npm package orderly-plugin-liquidity-heatmap (publisher dumtban, DEX "liquidity heatmap" costume) iterated 0.0.0-stage → 1.0.6 in a 3-hour window (13:11-16:05 UTC Oct 11) while its postinstall (scripts/setup.js) ran a full GitHub-Actions-runner reconnaissance battery: CI-gated snapshot of hostname/user/repo/workflow/GITHUB_* key inventory, hasGithubToken/tokenLen, OIDC mint probe through ACTIONS_ID_TOKEN_REQUEST_URL, live egress probes to four GitHub endpoints, tool-version inventory, then writes the snapshot INTO the victim's repository via the workflow's own GITHUB_TOKEN (PUT /repos/{repo}/contents/) and force-pushes a heatmap-setup-<ts> git tag impersonating github-actions[bot] (email 41898282+github-actions[bot]@users.noreply.github.com). The self-commentary frames every step as a proof-of-execution instrument ("proves both script execution and token reachability from a package install script", "unconditional trace … before any branching", pushing through GITHUB_STEP_SUMMARY "so it is visible via the public API even though yarn 1 swallows dependency-script stdout"), the author is demonstrating supply-chain reach, not hiding it. Zero registry action, zero OSV, zero GHSA; curated ossf/malicious-packages PR #1621 (opened 15:22:17Z Oct 11) is the only public record and is still OPEN. Its README cites a companion repo hashimubadiya14-rgb/asset-tracker-q4vw, same-persona test-harness grammar as the heatmap-setup-* tags it mints.

Artifact forensics (this wiki tarball pulls, no execution)

  • 0.0.0-stage 2026-10-11T13:11:40.439Z, then 1.0.0 13:12:56Z (76 s staging grammar), 1.0.1 13:53:05Z, 1.0.2 14:05:49Z, 1.0.3 14:20:01Z, 1.0.4 14:47:15Z, 1.0.6 16:05:29.590Z (1.0.5 absent, version churn mid-iteration). Maintainer: single account dumtban. latest = 1.0.6 LIVE at this check, installable, tarball 200.
  • Tarball (1.0.6, sha256 36c24c45…, 5 files): package.json declares scripts.postinstall: node ./scripts/setup.js; dist/index.js + README.md carry the costume (an "Orderly DEX liquidity heatmap plugin" with an OrderlyAppProvider registration snippet); the substance is scripts/setup.js.
  • The runner-profile snapshot (CI-gated: only fires with CI + GITHUB_WORKSPACE set): hostname, os.type/release, userInfo().username, cwd, GITHUB_REPOSITORY, GITHUB_WORKFLOW, RUNNER_ENV, ImageOS, the SORTED LIST of every GITHUB_*/ACTIONS_*/RUNNER_* environment KEY (names, not values), hasGithubToken + tokenLen, ACTIONS_ID_TOKEN_REQUEST_URL + presence of the request token, npm config get registry via execSync, curl status probes against registry.npmjs.org / api.github.com / github.com / objects.githubusercontent.com, and versions of git/node/npm/yarn/python3/docker/gh.
  • OIDC mint probe: if ACTIONS_ID_TOKEN_REQUEST_TOKEN + _URL are present it curls the runner's OIDC endpoint with &audience=api.github.com and records only the HTTP status (oidcMintProbe: "attempted"), self-commented as "audience probe only, token content is NOT read or transmitted, we only record whether minting succeeds". The probe still CONSUMES a real mint against the victim repo's identity, and it maps which federated-identity rails the runner exposes.
  • Local trace: appends heatmap-exec-trace.log to the temp dir (timestamp, pid, node version, cwd, CI/GITHUB flags) = an on-disk execution marker; on CI it also appends to GITHUB_STEP_SUMMARY.
  • The write-back (the durable harm): PUT https://api.github.com/repos/{GITHUB_REPOSITORY}/contents/heatmap-setup-report.json with Authorization: Bearer ${GITHUB_TOKEN}, commits the runner snapshot to the victim's default branch with message chore: liquidity-heatmap setup cache refresh [skip ci], reusing the file sha for overwrite. This turns any workflow that installs the package into a repo-writer: the token the checkout action grants (typically contents: write) is USED, not stolen.
  • The tag push (persistence-of-proof): git config to github-actions[bot] name + email 41898282+github-actions[bot]@users.noreply.github.com, then git tag -f heatmap-setup-<epoch-ms> and git push -f origin <tag> through the credential helper header actions/checkout writes into .git/config (http.extraheader). Result: attacker-influenced tags standing in the victim repo masquerading as the runner bot itself.

Read: proof-of-reach instrument, defender-relevant either way

The script's own comments argue this is an authorized-reach demonstration (everything is made publicly visible on purpose: step summaries, tags, committed reports, the opposite of exfil), and there is no secret-value theft in the artifact: key NAMES are inventoried, values are not read; the OIDC token is not captured. Three durable reads anyway:

  1. The capability map is a purchase-ready product. One npm install inside a victim CI job: reads runner topology, proves OIDC mintability, writes to the default branch with the workflow token, and mints bot-attributed tags. A malicious variant of this exact file changes two lines (PUT body → sensitive content; tag → persistence marker) and keeps every detection signature this benign build already passes, this is the living demonstration of why postinstall + GITHUB_TOKEN is a code-execution boundary, not a build detail.
  2. The iteration clock tracks the curated lane, not the registry: six versions in ~3 h, PR #1621 opened 15:22Z mid-iteration (after 1.0.4, before 1.0.6), the author kept shipping AFTER being publicly filed, and the registry did nothing: at this check the name is LIVE at flagged-adjacent 1.0.6, ZERO OSV + ZERO GHSA. This is the advisory-as-build-trigger costume shape (malfex, kafka-roller) measured on the CI-recon class.
  3. The bot-identity write is the quietest artifact. github-actions[bot] commits and force-pushed tags are invisible to humans who only watch PRs; detection needs heatmap-setup-* tag patterns, heatmap-setup-report.json path hits, and chore: liquidity-heatmap setup cache commit messages in repo/webhook logs.

IoCs

  • npm: orderly-plugin-liquidity-heatmap all versions (0.0.0-stage, 1.0.0-1.0.4, 1.0.6), publisher dumtban; tarball 1.0.6 sha256 36c24c456158bad255c5a6adc83247b02e2cda8561ebfe2860ccd1340472e2a2.
  • Files/paths: scripts/setup.js, heatmap-setup-report.json, heatmap-exec-trace.log (temp dir), step-summary header ## Liquidity Heatmap setup cache.
  • Git: tag pattern heatmap-setup-<epoch-ms> force-pushed with committer name github-actions[bot] / email 41898282+github-actions[bot]@users.noreply.github.com (the bot's REAL noreply address, the impersonation is in the actor name, indistinguishable from a real runner write except by provenance).
  • Commit message: chore: liquidity-heatmap setup cache refresh [skip ci]; User-Agent liquidity-heatmap-setup.
  • Companion repo cited in README: hashimubadiya14-rgb/asset-tracker-q4vw (persona + <slug>-<4rand> grammar shared with heatmap-setup naming).

Status & monitor

  • At 23:3x-23:5x UTC Oct 11: name LIVE latest 1.0.6, zero OSV (v1/query {}), zero GHSA, no registry action; PR ossf/malicious-packages #1621 OPEN (the only record); companion repo not probed beyond the README reference.
  • OCT 12 one-hundred-and-fifty-third sweep re-check (~01:3x UTC): latest 1.0.6 UNCHANGED ≈9.5 h after PR #1621, no 1.0.7, modified frozen at 16:05:29Z, zero OSV ({}) + zero GHSA, registry still inert, post-PR iteration-or-silence clock open; the silence-so-far is itself data vs the malfex/kafka-roller same-day re-fire shape.

OCT 13 one-hundred-and-seventeenth sweep (~07:2x-07:4x UTC): THE ADVISORY CHAIN LANDS, QUEUE-MERGE→GHSA IN 2 m 40 s, REGISTRY STILL AT ZERO ≈1 h LATER

  • #1621 MERGED by maintainer calebbrown at 2026-10-13T06:27:48Z, ≈39 h after filing. Within the same import cycle: MAL-2026-17848 LIVE (modified 06:30:05.184Z) and GHSA-fp2f-2254-46g7 critical published 06:30:28Z, listing head moved. Scope: six EXPLICIT version entries (= 1.0.0-= 1.0.4, = 1.0.6), hand-scoped, no >=0 boilerplate; 0.0.0-stage NOT covered (staging-stub blind spot reproduced on the curated lane). aliases still None ≈1 h post-mint (forward-key/cache read, standing rule).
  • THE PUBLISHED FIELD IS BACKDATED: the OSV record reads published 2026-10-11T13:13:56Z, ≈72 h BEFORE the merge that minted the id (it sits between the package's own 1.0.0 13:12:56Z and 1.0.1 13:53:05Z publishes). Same backfill grammar as the 148th's RubyGems fleet. DURABLE: published on the curated lane is the queued-PR-era claim; the modified flip is the mint event. Any exposure clock computed from published on a curated-lane record is measuring the queue, not the attacker.
  • THE 42-HOUR FALSE-CLEAN WINDOW IS NOW A MEASURED FACT: from PR filing (Oct 11 15:22Z) to advisory (Oct 13 06:30Z), every OSV/GHSA-keyed consumer, Dependabot included, read CLEAN on a name the curated queue, this ledger, and the record's own summary text all described as malicious. The advisory text is word-for-word the analysis that sat in the OPEN PR for two days. Queue latency is exposure, end to end.
  • Registry disposition at ≈1 h 05 m post-GHSA: ZERO. npm doc UNCHANGED (latest 1.0.6, modified frozen 2026-10-11T16:05:29Z), tarball 200/3,321 B installable, no holder, no unpublish, 0.0.0-stage live. Advise-then-installable reproduces on the curated-merge lane exactly as on the kam193 direct lane (npp-ml, ≈11 h and counting). The advisory is the consumer-visible layer; enforcement is not on any guaranteed clock.
  • Watch dispositions: zero-record watch CLOSED (advisory arrived); post-#1621 iteration watch CLOSED with a NEGATIVE result (no 1.0.7 ever shipped, six versions total, the author went silent after 1.0.6); NEW watches open: npm disposition, dumtban account activity, re-registration/re-arm per the re-publish-triggered enforcement rule, heatmap-setup-* tag sightings, 17848 alias join (does a curated-merge-minted record mirror its GHSA alias faster than the direct-analyst lane? the 115th's tail-not-band question gets a second subject).
  • Monitor: 1.0.7+ or any post-#1621 iteration (build-trigger confirmation); npm abuse-lane disposition (delete vs 0.0.1-security holder vs nothing); OSV/GHSA arrival clock vs the curated PR; re-registration under the same publisher; whether the runner-profile→repo-write pattern is adopted by other names (hunt: postinstall scripts that reference GITHUB_STEP_SUMMARY, ACTIONS_ID_TOKEN_REQUEST_URL, or PUT /contents/); heatmap-setup-* tag sightings in public repos via fork/gitleaks-class crawls; dumtban account recovery/action.