Supply-chain threat intelligence

Incident detail

criticalpypi·maintainer compromise·osv

Malicious code in mrmustard (PyPI)

mrmustard

Risk score

92

AI summary

Indexed incident for mrmustard (pypi).

Description

On plain import mrmustard, the top-level init.py spawns a background thread that reads ~/.ssh/ private keys, ~/.aws/credentials, ~/.aws/config, ~/.kube/config, environment variables, pip freeze output, and host/GPU/SLURM identifiers. The collected data is XOR-encoded with the key 'tf_compat_v2', base64-wrapped, and POSTed via urllib to an endpoint reconstructed at runtime from an obfuscated base64+XOR literal, using a spoofed browser User-Agent. Before firing, the code returns early when CI, GITHUB_ACTIONS, GITLAB_CI, JENKINS_URL, TRAVIS, CIRCLECI, BUILDKITE, or CODEBUILD_BUILD_ID is set, when /.dockerenv or /run/.containerenv exists, or when /proc/1/cgroup mentions docker/kubepods/lxc — so the payload only executes on developer/production hosts. Three persistence mechanisms are installed under the cover story of a 'tensorflow hardware compatibility check': a compiled dropper is written to ~/.cache/.tf_cache/hw_probe.pyc, a crontab entry runs it every 15 minutes, an 'mmcompat.pth' file is dropped into site-packages so it executes on every Python startup, and a launcher line is appended to ~/.bashrc, ~/.zshrc, and the fish shell config. These mechanisms continue to run the exfiltration payload after the package is uninstalled. The legitimate MrMustard (Xanadu) package does not exhibit this behavior; this version is a compromised or impersonating release.

Versions 0.7.4 were compromised.

Compromised release has embedded code that during import exfiltrates sensitive data (selected environmental variables, credentials to AWS, SSH keys etc.) and ensures persistence via multiple ways: a cron entry, a malicious PTH file, and a shell configuration file. Persistence is diguished as "tensorflow hardware compatibility check" using file placed under ~/.cache/.tf_cache/hw_probe.pyc. The malicious version was uploaded after exfiltrating the PyPI token from the CI environment, likely after compromising the maintainer's Github account.


Category: MALICIOUS - The campaign has clearly malicious intent, like infostealers.

Campaign: 2026-07-compr-hw-probe

Reasons (based on the campaign):

  • exfiltration-env-variables

  • exfiltration-ssh-keys

  • The package contains code to detect if it is running in a sandbox environment.

  • exfiltration-credentials

  • persistence

  • compromised-package

  • abuses-pth

Technical details

Affected versions

=0.7.4

Indicators

  • affected version=0.7.475%

Timeline

  1. Advisory published
  2. Indexed by ThreatPkg

Related incidents