[Digital Today reporter Chi-gyu Hwang] LiteLLM, a tool for connecting AI models, had its deployment system taken over by hackers after an administrator key was mistakenly made public, The Hacker News reported on Oct. 24.

The report said LiteLLM disclosed the circumstances of the incident and said so.

LiteLLM said a configuration file uploaded to a public GitHub repository on May 27 contained an agent platform master key. The company quickly realised the mistake and deleted the file, but did not revoke the key.

A hacker entered the agent platform with the key on Oct. 10 and stole a GitHub access token. The token was tied to a top administrator account with no permission restrictions.

The hacker accessed all repositories in the organisation and stole tokens from 2 engineers with administrative privileges.

The hacker changed the deployment path to a fake repository on Oct. 21 and started deployment with one of the engineer tokens, then approved it with the other. A malicious version, 1.107.0, was then uploaded to the Python package repository PyPI and to Docker Hub, among other places. A researcher at Upwind reported it in less than 30 minutes, and PyPI quarantined the version about an hour later.

Microsoft's security research team warned that the malicious file was outside the app folder and could remain if only the app files were deleted.