Skip to main content
Back
IntrusionAWS · Azure · Microsoft Entra · Google Cloud · Kubernetes · SaaS
9 min read

durabletask PyPI releases carry a stealer attributed to TeamPCP

Victim
Microsoft (the durabletask Python client project) and the users who imported versions 1.4.1 to 1.4.3
When
Token theft date not stated; staging from May 16 – May 20, 2026Disclosed May 19, 2026
Entry
The project's long-lived PyPI API token, stored as a GitHub Actions secret, was used to upload three releases from outside the release workflow, by the research write-ups' account; the advisories name a compromised maintainer account, and how the credential was obtained is unconfirmed.

On May 19, 2026, durabletask 1.4.1, 1.4.2 and 1.4.3 were uploaded to PyPI within 35 minutes with no tag and no workflow run behind them; the research write-ups attribute the uploads to the project's long-lived PyPI token, the advisories to a compromised maintainer account, and how the credential was obtained is unconfirmed. On Linux, importing a release is built to download a second stage that is written to read cloud credentials and secret stores on AWS, Azure and Google Cloud, read every Kubernetes Secret it can reach, run itself on up to five more instances and five more pods, and post an encrypted harvest out. Microsoft yanked the releases the same day and PyPI quarantined the project on May 20; no source names an affected organization or confirms that credentials were taken or used.

What happened

External Parties
Command servers
Victims' GitHub accounts
Researchers
Internet
Instance Metadata ServiceAWS, Azure, Google Cloud
GitHub
durabletask-python Repository
Environments That Imported It
+ durabletask project
+ releases 1.4.1 to 1.4.3
+ installs from the index
PyPI
PyPI Package Index
Actions Secrets
Release Workflow
Linux Environment
Managed Instances
Cloud Secret Stores
Kubernetes API + Pods

Replaying the durabletask May 2026 intrusion

  1. 1

    Publishing token

    Token theft not stated; uploads May 19, 2026, 16:19–16:54 UTC

    A credential for the PyPI project, by the research write-ups the token kept as a repository secret, uploads three releases outside the release workflow; how it was obtained is unconfirmed.

  2. 2

    Import and run

    From May 19, 2026, 16:19 UTC

    On Linux, importing a release is built to download a second stage and start it as a detached process.

  3. 3

    Cloud secrets and spread

    From May 19, 2026, 16:19 UTC

    The payload is written to ask the metadata services for workload credentials, read every secret store they reach and run itself on other instances through Systems Manager; no source confirms it did so in a named account.

  4. 4

    Cluster spread

    From May 19, 2026, 16:19 UTC

    With a kubeconfig or an in-cluster token, the payload is written to read every Secret of every namespace and run itself in up to five other pods through exec; no source confirms it did so in a named cluster.

  5. 5

    Exfiltration and cleanup

    May 19–20, 2026

    The payload is built to post its encrypted harvest out, fall back to public repositories made with stolen GitHub tokens and, on the operator's switch, install a disguised service; researchers flag the releases and PyPI quarantines the project.

Impact

Three malicious durabletask releases were uploaded to PyPI on May 19 within 35 minutes and yanked the same day. In each Linux environment that imported one, the payload was built to send its cloud, Kubernetes, Vault and password-manager secrets to the attacker and to spread through Systems Manager and pod exec to five more hosts each. No source names an affected organization or confirms that credentials were taken or used.

Outcome

Where cloud-provider controls fit

The publishing credential, the upload and the import sit with GitHub, PyPI and the package manager, so no cloud control reaches them: environment rules on the repository secret, PyPI trusted publishing and pinned, hashed dependencies are yours to set there. Had the payload run in your environment, cloud controls would bound what it took: metadata settings that keep pods from the node's credentials, roles and network conditions that limit which secrets a credential reads, and policies that keep remote commands to your administrators. In a cluster, a private API endpoint, a pod with no mounted token and an admission policy on exec would bound its spread, where no cloud control reaches a pod exec the cluster allows. An egress allow list on AWS, Azure or Google Cloud, or a NetworkPolicy in the cluster, would bound the POST that sends the harvest out. The payload would still read what its own host may read: you own the rotation of every credential an importing host could reach.

Could this have been prevented?

In hindsight, no control would have stopped it. 22 controls would have limited how far it went. 7 steps have no cloud-provider control. The counts below are control assessments, not attack stages: several can address the same step. The same replay with the controls in place before the incident shows where each phase would have been refused or bounded.

0
Prevented
22
Contained
7
Not covered

How it could have been prevented

The same replay with the controls in place. Green marks where a step is refused, blue where it is bounded, and red where no control reaches it. Open a row to read the control, why it acts, and its caveat.

External Parties
Command servers
Victims' GitHub accounts
Researchers
Internet
Instance Metadata ServiceAWS, Azure, Google Cloud
GitHub
durabletask-python Repository
Environments That Imported It
+ durabletask project
+ releases 1.4.1 to 1.4.3
+ installs from the index
PyPI
PyPI Package Index
Actions Secrets
Release Workflow
Linux Environment
Managed Instances
Cloud Secret Stores
Kubernetes API + Pods

Replaying the durabletask May 2026 intrusion

1 · Publishing token

1 Contained2 Not covered

2 · Import and run

1 Not covered

3 · Cloud secrets and spread

11 Contained3 on other clouds

4 · Cluster spread

5 Contained1 Not covered1 on other clouds

5 · Exfiltration and cleanup

5 Contained3 Not covered1 on other clouds

Key facts

  • durabletask 1.4.1, 1.4.2 and 1.4.3 were uploaded to PyPI on May 19, 2026, at 16:19, 16:49 and 16:54 UTC, with no tag and no workflow run behind them, per Upwind, SafeDep, StepSecurity and the OSV record
  • The research write-ups attributed the uploads to a long-lived PyPI API token kept as a GitHub Actions secret, while the advisories named a compromised maintainer account; Wiz inferred that a GitHub account compromised in an earlier wave reached the repository secrets, SafeDep said the repository was not breached, and how the credential was obtained is unconfirmed
  • On Linux, importing a release was built to download a second stage and start it as a detached process, skipping hosts with a Russian locale or with two CPUs or fewer, per Upwind and the OSV record
  • The payload's code was written to read secrets across 19 AWS regions, every key vault and Secret Manager, and to run itself on up to five managed instances and five pods per run, per SafeDep, Upwind and Endor Labs; no source confirms that it ran in a named environment
  • Endor Labs flagged 1.4.1 at 16:21 UTC, a Microsoft maintainer wrote on May 19 that the releases were yanked, and on May 20 PyPI quarantined the project and suspended the associated maintainer accounts
  • The package had about 417k monthly downloads, which Endor Labs gives as exposure, not a victim count; Wiz linked the campaign to TeamPCP, PyPA called that likely and Aikido could not be sure

Sources

Every fact in this breakdown comes from the public reporting below. The replay is a model of that reporting, not a reconstruction of the environment it describes.

  1. 1.
    Potential malicious code execution in durabletask versions 1.4.1, 1.4.2, and 1.4.3 (issue #137)Microsoft (durabletask-python maintainers) · May 19, 2026 · Vendor reportresponse, yank, quarantine, maintainer account suspension, release pipeline change, trusted publishing for 1.5.0
  2. 2.
    Newly Discovered durabletask Malware Targeted Kubernetes, Cloud Secrets, and CI/CD InfrastructureUpwind Research · May 19, 2026 · Researchupload times, publishing token, release-flow evidence, payload trigger, AWS, Azure and GCP harvest, SSM propagation, Kubernetes behavior, persistence, exfiltration, wiper, staging timeline
  3. 3.
    Malicious durabletask on PyPI: Multi-Cloud Credential Stealer with Worm CapabilitiesSafeDep · May 20, 2026 · Researchpublishing token, build path, injection points, AWS, Azure and GCP harvest, SSM propagation, Kubernetes behavior, GitHub exfiltration fallback, persistence, wiper
  4. 4.
    Microsoft's durabletask PyPI Package Compromised in Supply Chain AttackStepSecurity · Jun 5, 2026 · Researchrelease workflow, no tag and no workflow run, upload times, runtime observations, SSM and Kubernetes behavior, follow-on repository commit
  5. 5.
    durabletask: TeamPCP's Latest PyPi CompromiseWiz Research · May 19, 2026 · ResearchGitHub account inference, repository activity window, publishing token, injection points, capability table, attribution, quarantine
  6. 6.
    Trojanized Microsoft SDK durabletask 1.4.1 through 1.4.3 Deliver Credential-Stealing MalwareEndor Labs · May 19, 2026 · Researchdetection time, payload trigger, SSM propagation, Kubernetes propagation, GitHub exfiltration fallback, download volume
  7. 7.
    durabletask package compromised (Mini Shai-Hulud)Aikido Security · Sep 24, 2026 · Researchpayload trigger, AWS, Azure and GCP harvest, SSM propagation, Kubernetes behavior, exfiltration, persistence, wiper, attribution hedge
  8. 8.
    MAL-2026-4174: Malicious code in durabletask (PyPI)OpenSSF Malicious Packages (OSV) · Jun 10, 2026 · Advisorypublication window, import-time trigger, second-stage capture, payload summary, maintainer account wording
  9. 9.
    PYSEC-2026-207: durabletask 1.4.1, 1.4.2, and 1.4.3 contain malicious code distributed via a compromised maintainer accountPython Packaging Advisory Database (PyPA) · Jun 28, 2026 · Advisorymaintainer account wording, removal from PyPI, payload summary, attribution as likely, 1.5.0 release
  10. 10.
    Switch to Trusted Publishing for package upload to PyPI in CI (issue #139)GitHub issue by an outside contributor (microsoft/durabletask-python) · May 20, 2026 · Researchrelease workflow authentication, stored PyPI token

The estate in the replay is illustrative and drawn at the level of attack stages; the victim's real environment is not public. Verdicts describe what a control does when it is in place before the incident. They are not an assessment of the victim's security program.

More cloud attack breakdowns

All breakdowns