PyPI

The num2words PyPI Supply Chain Attack

Updated 5 Oct 2026 · Incident date 28 Jul 2025

num2words 0.5.15 and 0.5.16 on PyPI held malware that stole files and gave remote access, uploaded with a phished maintainer login on 28 July 2025.

Version 0.5.14 was clean. The code ran when the package was imported on Windows and loaded a native DLL named Scavenger.

What happened

A phishing email sent the package owner to a fake PyPI site at pypj.org, which used a lowercase j in place of the i.

  • The PyPI incident report says four accounts were phished and two API tokens were created, then revoked.
  • The attacker used the stolen credentials to upload the bad releases.
  • On Windows, the bad __init__.py loads _build.py, which is a DLL, and calls its main export through ctypes. This is the Scavenger loader.
  • The loader can fetch stealer modules: "shiny" as version.dll, "electric" as umpdc.dll and "exodus" as profapi.dll.
  • The malware goes after .pypirc files, wallet files, and browser and system credentials.

The upload came on 28 July 2025. The registrar NameSilo placed the phishing domain on administrative hold the same day. The advisory is PYSEC-2025-72.

Affected versions

  • 0.5.15: malicious, contains the loader DLL.
  • 0.5.16: a re-release that still contained the backdoor, according to Invoke RE.
  • 0.5.14 was the last clean release.

Indicators of compromise

  • Loader SHA-256: be917cb379b9622f56a4d5ec93bf00c20cb76c6646e5919690d0f7c09c956de2.
  • Stealer "electric" SHA-256: c2a7ee6ab9344e1bb13c61dc689d4a946678e0505367cd55c9b43ddee3d461e2.
  • Stealer "exodus" SHA-256: 439da8bb9c541d26b0f534b17d75790e252e4d9058561e8907f8690e21cd0616.
  • Stealer "shiny" SHA-256: c36ebf96573afcb36bb31590d56e8af49502fb159e00fd4a59336f8a450bec8b.
  • Command domains: ifyouseethisyouareultragay.com and pokerainteasy.su.
  • Phishing domain: pypj.org, hosted at 45.9.148.85.
  • A _build.py file inside the installed num2words folder.

How to check

Check the installed version. If the result is 0.5.15 or 0.5.16, the bad package is installed.

pip show num2words

Look for the loader file in the package folder. The command does not import the package.

pip show -f num2words | grep -i _build

What to do now

  1. Uninstall the bad version and install a clean one, such as 0.5.14 or a later release you have checked.
  2. If a bad version was imported on Windows, treat the machine as compromised.
  3. Rotate credentials in any .pypirc file, plus browser and system credentials stored on the machine. Move wallet funds if wallet software was present.
  4. Block the command domains in the IOC list.

What Vigilance would show

These examples show an install hold and the corresponding owner report.

Example install message

Vigilance blocked `num2words` 0.5.15. This version is too new or changed in a way Vigilance must check first.

Example owner report

Bad update

num2words 0.5.15

A number-spelling library gained a hidden bundled program that stole files and gave outsiders lasting remote access to the machine.

See what to do now.

No security team? See who reads the alert.

Frequently asked questions

Does the loader run on Linux and macOS?

No. The loader code runs only on Windows. Linux and macOS systems still received the bad package files.

How can publishers protect PyPI accounts?

PyPI recommends WebAuthn two-factor sign-in with a hardware key, and deleting dormant accounts.

Sources

  1. blog.pypi.org/posts/2025-07-31-incident-report-phishing-attack/
  2. invokere.com/posts/2025/07/scavenger-malware-distributed-via-num2words-pypi-supply-chain-compromise/
  3. github.com/Invoke-RE/community-malware-research/blob/main/Research/Loaders/Scavenger/num2words_IOCs.md

Send this page to the person who looks after AI at your company.

Email it

More supply chain attacks

All 172 attacks in the library · What is a supply chain attack? · How to prevent supply chain attacks

Check the next update before you install it

Vigilance compares the version you trust with the new one. It names the one file that gained a power the old version did not have.

Start the free 14-day trialBrowse the attack library