The Malicious npm Release Had Valid Provenance Because the Build System Worked
The poisoned package was not smuggled around the build system. GitHub Actions built it, npm trusted the workflow identity and Sigstore recorded the provenance. Every control accurately described a release created from an attacker’s commit.
CloudSEK’s public investigation calls the loader operation GHAPPIER. Its investigation began with version 0.2.21 of the legitimate @dforge-core/dforge-mcp package and expanded to at least 65 public repositories, 73 infected files and 22 accounts.
Push access became publish access
The intruder controlled the maintainer’s repository for 105 minutes on 9 September. They inserted a one-line remote loader, changed the workflow so pushes to the main branch triggered releases and published the malicious version without needing a separate npm token. How the initial push access was obtained remains unexplained. CloudSEK suspects stolen developer credentials from a compromised machine, but does not establish that initial route as fact.
Trusted publishing did what it was configured to do. It accepted the repository’s CI identity. The resulting attestation points to the attacker’s real commit, which is valuable forensic evidence but not a verdict that the code was safe. npm’s provenance documentation makes that limit explicit: provenance links a package to its source and build instructions, rather than certifying the absence of malicious code.
The package remained the latest release for 35 minutes and 38 seconds before the maintainer reverted the changes and published 0.2.22. CloudSEK says the later stages remained reachable five days after withdrawal.
The final implant removed itself
The first malicious line started a four-stage chain. The final implant deleted itself from disk at startup, so an organisation that removed the package and searched only for the last payload could reasonably reach the wrong conclusion.
CloudSEK also found a related payload pattern associated with the PolinRider campaign. Some attribution claims in the wider reporting point towards North Korea, but CloudSEK says its own independent check did not confirm that final attribution step. The documented supply-chain activity stands without it.
What maintainers and consumers should change
- Protect the release boundary. Require protected branches, independent review and restricted workflow-file changes for release repositories. GitHub’s secure-use guidance recommends reviewing workflow files with designated owners and limiting workflow-token permissions.
- Review who can change the publisher. Treat GitHub push authority as production publishing authority where the configured OIDC trusted-publishing workflow permits a release. npm’s trusted-publisher guidance uses short-lived workflow identities rather than a reusable registry secret. Review the authorised repository, workflow and environment as well as the source; removing a token does not remove those permissions.
- Fix the dependency and investigate execution. Pin or upgrade the affected package to the clean 0.2.22 identified by the researchers, and investigate any execution of 0.2.21. Withdrawal alone does not clear a machine that already ran the loader.
- Use the evidence rather than discarding it. Use the provenance record to identify the exact malicious commit and affected build, not as a substitute for source review. SLSA’s provenance model distinguishes tracing a build’s output to its source from tracing the source revision and its change-management process.
- Hunt across the whole sequence. Hunt for intermediate artefacts and network behaviour because the final implant deletes itself. Preserve evidence before cleaning; a negative search for the final filename is insufficient.
- Contain the wider development access. Revoke exposed developer credentials and review other repositories reachable from the same workstation or account. The NCSC’s incident-response guidance stresses establishing scope and preserving evidence rather than assuming the first visible asset defines the incident.
CloudSEK says it found no evidence that an organisation was successfully compromised through this release. The incident still exposes a dangerous misunderstanding: provenance can prove that a trusted factory produced the package while simultaneously proving that the factory received poisoned instructions.
Sources
- CloudSEK GHAPPIER investigation, page dated 20 September 2026. Its machine-readable publication field says 21 September, while the modified field says 20 September, so those values do not establish a consistent exact publication time.


