AI
Malicious LiteLLM 1.107.0 on PyPI steals credentials, SafeDep reports
One version of a popular AI toolkit is reported to run credential-stealing code every time Python starts.
Security firm SafeDep reports that LiteLLM 1.107.0, a version of a popular Python toolkit linking programs to many AI providers, was published to PyPI (Python's public package library) with credential-stealing code. SafeDep says 1.106.2 and the 1.107.0 release candidates are clean.
According to SafeDep, a file called cook.pth runs on every Python start and targets SSH keys, cloud credentials, CI/CD secrets (passwords for automated build systems), browser passwords and AI provider keys, uploading them to a site it names. A second piece, sitecustomize.py, runs an hourly background task that executes downloaded code. SafeDep's indicators include those files in site-packages (where Python keeps installed packages) and requests to 169.254.169.254.
The caveat: this is SafeDep's own analysis, not independently confirmed, and the article has no remediation section.
On your desk: run `pip show litellm` in any environment behind your websites or AI tools. If it says 1.107.0, rotating the keys on SafeDep's target list is our inference, not its instruction.
Read the original at SafeDep: LiteLLM PyPI compromise, credential stealer and backdoor →