Postmortem: TanStack npm supply-chain compromise

by Tanner Linsley on May 11, 2026.

Last updated: 2026-05-11

TL;DR#

On 2026-05-11 between 19:20 and 19:26 UTC, an attacker published 84 malicious versions across 42 @tanstack/* npm packages by combining: the pull_request_target "Pwn Request" pattern, GitHub Actions cache poisoning across the fork↔base trust boundary, and runtime memory extraction of an OIDC token from the GitHub Actions runner process. No npm tokens were stolen and the npm publish workflow itself was not compromised.

The malicious versions were detected publicly within 20 minutes by an external researcher ashishkurmi working for stepsecurity. All affected versions have been deprecated; npm security has been engaged to pull tarballs from the registry. We have no evidence of npm credentials being stolen, but we strongly recommend that anyone who installed an affected version on 2026-05-11 rotate AWS, GCP, Kubernetes, Vault, GitHub, npm, and SSH credentials reachable from the install host.

Tracking issue: TanStack/router#7383 GitHub Security Advisory: GHSA-g7cv-rxg3-hmpx

Impact#

Packages affected#

42 packages, 84 versions (two per package, published roughly 6 minutes apart). See the tracking issue for the full table. Confirmed-clean families: @tanstack/query*, @tanstack/table*, @tanstack/form*, @tanstack/virtual*, @tanstack/store, @tanstack/start (the meta-package, not @tanstack/start-*).

What the malware does#

When a developer or CI environment runs npm install, pnpm install, or yarn install against any affected version, npm resolves the malicious optionalDependencies entry, fetches the orphan payload commit from the fork network, runs its prepare lifecycle script, and executes a ~2.3 MB obfuscated router_init.js smuggled into the affected tarball. The script:

• Harvests credentials from common locations: AWS IMDS / Secrets Manager, GCP metadata, Kubernetes service-account tokens, Vault tokens, ~/.npmrc, GitHub tokens (env, gh CLI, .git-credentials), SSH private keys
• Exfiltrates over the Session/Oxen messenger file-upload network (filev2.getsession.org, seed{1,2,3}.getsession.org) — end-to-end encrypted with no attacker-controlled C2, so blocking by IP/domain is the only network mitigation
• Self-propagates: enumerates other packages the victim maintains via registry.npmjs.org/-/v1/search?text=maintainer:<user> and republishes them with the same injection

Because the payload runs as part of npm install's lifecycle, anyone who installed an affected version on 2026-05-11 must treat the install host as potentially compromised.

Timeline#

All times UTC. Local timestamps from GitHub API and npm registry.

Pre-attack (cache poisoning phase)#

Detonation (publish phase)#

Detection and response#

Root cause#

Three vulnerabilities chained together. Each is necessary for the attack; none alone is sufficient.

1. pull_request_target "Pwn Request" pattern in bundle-size.yml#

bundle-size.yml ran pull_request_target for fork PRs and, inside that trigger context, checked out the fork's PR-merge ref and ran a build:

on:
  pull_request_target:
    paths: ['packages/**', 'benchmarks/**']

jobs:
  benchmark-pr:
    steps:
      - uses: actions/checkout@v6.0.2
        with:
          ref: refs/pull/${{ github.event.pull_request.number }}/merge # fork's merged code

      - uses: TanStack/config/.github/setup@main # transitively calls actions/cache@v5

      - run: pnpm nx run @benchmarks/bundle-size:build # executes fork-controlled code

on:
  pull_request_target:
    paths: ['packages/**', 'benchmarks/**']

jobs:
  benchmark-pr:
    steps:
      - uses: actions/checkout@v6.0.2
        with:
          ref: refs/pull/${{ github.event.pull_request.number }}/merge # fork's merged code

      - uses: TanStack/config/.github/setup@main # transitively calls actions/cache@v5

      - run: pnpm nx run @benchmarks/bundle-size:build # executes fork-controlled code

The author of the workflow attempted a trust split (the comment-pr job is separate from benchmark-pr, with a comment in the YAML noting the intent to keep benchmark-pr "untrusted with read-only permissions"). The split is correct in spirit but missed two facts:

• actions/cache@v5's post-job save is not gated by permissions:. Cache writes use a runner-internal token, not the workflow GITHUB_TOKEN. Setting permissions: contents: read does not block cache mutation.
• Cache scope is per-repo, shared across pull_request_target runs (which use the base repo's cache scope) and pushes to main. A PR running in the base repo's cache scope can poison entries that production workflows on main will later restore.

2. GitHub Actions cache poisoning across trust boundaries#

The malicious vite_setup.mjs was specifically designed to write data into the pnpm-store directory under a key the legit release.yml workflow would compute and look up: Linux-pnpm-store-${hashFiles('**/pnpm-lock.yaml')}. When the benchmark-pr job ended, actions/cache@v5's post-step saved the (now-poisoned) pnpm store to that exact key. When release.yml next ran on a push to main, its Setup Tools step restored the poisoned entry — entirely as designed.

This is the class of attack documented by Adnan Khan in 2024. It's not a TanStack-specific bug; it's a known GitHub Actions design issue that requires conscious mitigation.

3. OIDC token extraction from runner memory#

release.yml declares id-token: write (legitimately needed for npm OIDC trusted publishing). When the poisoned pnpm store is restored on the runner, attacker-controlled binaries are now on disk and get invoked during the build step. Those binaries:

• Locate the GitHub Actions Runner.Worker process via /proc/*/cmdline
• Read /proc/<pid>/maps and /proc/<pid>/mem to dump the worker's memory
• Extract the OIDC token (which the runner mints lazily, in memory, when id-token: write is set)
• Use the token to authenticate POST requests directly to registry.npmjs.org — bypassing the workflow's Publish Packages step entirely

This is the same memory-extraction technique (and verbatim Python script, with attribution comment) used in the tj-actions/changed-files compromise of March 2025. The attacker did not invent novel tradecraft; they recombined published research.

Why none alone is enough#

• pull_request_target alone is fine for trusted operations (labeling, comments)
• Cache poisoning alone (e.g., from inside an already-compromised dep) requires a separate publish vehicle
• OIDC token extraction alone requires existing code execution on the runner

The chain only works because each vulnerability bridges the trust boundary the others assumed: PR fork code crossing into base-repo cache, base-repo cache crossing into release-workflow runtime, and release-workflow runtime crossing into npm registry write access.

Detection#

How we found out#

Detection was external. carlini opened issue #7383 ~20 minutes after the publish, with full technical analysis. Tanner received a phone call from Socket.dev just moments after starting the war room confirming the situation.

IOC fingerprints (for downstream maintainers and security tools)#

In any @tanstack/* package's manifest:

"optionalDependencies": {
  "@tanstack/setup": "github:tanstack/router#79ac49eedf774dd4b0cfa308722bc463cfe5885c"
}

"optionalDependencies": {
  "@tanstack/setup": "github:tanstack/router#79ac49eedf774dd4b0cfa308722bc463cfe5885c"
}

• File: router_init.js (~2.3 MB, package root, not in "files")
• Cache key: Linux-pnpm-store-6f9233a50def742c09fde54f56553d6b449a535adf87d4083690539f49ae4da11
• 2nd-stage payload URLs: https://litter.catbox.moe/h8nc9u.js, https://litter.catbox.moe/7rrc6l.mjs
• Exfiltration network: filev2.getsession.org, seed{1,2,3}.getsession.org
• Forged commit identity: claude <claude@users.noreply.github.com> (note: not the real Anthropic Claude — fabricated GitHub no-reply email)
• Real attacker accounts: zblgg (id 127806521), voicproducoes (id 269549300)
• Attacker fork: github.com/zblgg/configuration (fork of TanStack/router renamed to evade fork searches)
• Orphan payload commit (in fork network): 79ac49eedf774dd4b0cfa308722bc463cfe5885c
• Workflow runs that performed the malicious publishes:
  • github.com/TanStack/router/actions/runs/25613093674 (attempt 4)
  • github.com/TanStack/router/actions/runs/25691781302

Lessons learned#

What went well#

• External researchers noticed and reported with full technical detail within ~20 min of the incident
• Maintainer team coordinated immediately and effectively across many timezones
• The detection community already had a clear public IOC pattern within hours

What could have been better#

• No internal alerting. We learned about the compromise from a third party. We need monitoring on our own publishes. We'll be working closely with security researcher firms in the ecosystem that have the ability to detect these issues very quickly, potentially even in-house, and making the feedback loop even tighter.
• pull_request_target workflows had not been audited despite being a long-known dangerous pattern
• Floating refs (@v6.0.2, @main) on third-party actions create standing supply-chain risk independent of this incident
• Unpublish was unavailable for nearly all affected packages because of npm's "no unpublish if dependents exist" policy. We have to rely on npm security to pull tarballs server-side, which adds hours of delay during which malicious tarballs remain installable
• The 7-maintainer list on the npm scope means seven separate credential-theft targets for the same blast radius
• OIDC trusted-publisher binding has no per-publish review. Once configured, any code path in the workflow can mint a publish-capable token. We need either (a) move to short-lived classic tokens with manual review, or (b) add provenance-source-verification to detect publishes from unexpected workflow steps

What we got lucky on#

• The attacker chose a payload that broke tests, which made the publish step (which would have produced cleaner-looking tarballs) skip — meaning the attack was loud enough to detect quickly. A more careful attacker who didn't break tests could have published silently for hours longer
• The attacker reused public tradecraft (verbatim memory-dump script with attribution comment) instead of writing novel code — making the IOC-matching faster

Open questions#

These need answers before we close the postmortem.

• Did bundle-size.yml's Setup Tools step actually call actions/cache@v5? Verify by reading the post-job logs from one of the pull_request_target runs against PR #7378 (e.g., run id 25666610798). Tanner has access; needs to be done manually
• What was in the initial PR head commit (before the force-pushes wiped it)? GitHub's reflog may have it. Check via gh api or the GitHub support team
• How did the malicious commit get into the fork's git object store specifically — was it pushed directly via git, or was it created via the GitHub web UI (which would leave audit-log entries)?
• Was voicproducoes a real account or a sock puppet? Cross-reference its activity history
• Did the npm cache also get poisoned (the 6 duplicate linux-npm-store-* entries)? Were any actually used?
• Does the attack require Nx Cloud, or would it have worked with just GitHub Actions cache?
• Can we identify any other fork in the TanStack/router fork network that contains the orphan payload commit? (If yes, the cleanup is harder — every fork hosting it keeps it accessible via github:tanstack/router#79ac49ee...)
• Are any other TanStack repos (router, query, table, form, virtual, etc.) using the same bundle-size.yml-style pattern? Audit needed
• How many users actually downloaded the affected versions during the publish window? Get from npm support
• Did any of the seven listed maintainers' machines get compromised separately? (None of the malicious publishes used a maintainer's npm token, but maintainer machines could have been the secondary target via the self-propagation logic)

References#

• Tracking issue: TanStack/router#7383
• GitHub Security Advisory: GHSA-g7cv-rxg3-hmpx
• Related research:
  • Adnan Khan, "The Monsters in Your Build Cache: Github Actions Cache Poisoning" (May 2024) — adnanthekhan.com
  • GitHub Security Lab, "Keeping your GitHub Actions and workflows secure: Preventing pwn requests" — securitylab.github.com
  • StepSecurity, "Harden-Runner detection: tj-actions/changed-files action is compromised" (March 2025) — stepsecurity.io
• npm policies:
  • Unpublish: docs.npmjs.com/policies/unpublish
  • Provenance: docs.npmjs.com/generating-provenance-statements

Appendix A — Affected versions#

See the GitHub Security Advisory for the full list of affected versions: GHSA-g7cv-rxg3-hmpx