LiteLLM Hijacked: Malicious 1.107.0 Wheels Steal Credentials (research.jfrog.com)

The JFrog Security Research team has discovered that version 1.107.0 of LiteLLM, a popular Python interface for large language models, contains malicious code that steals credentials and allows for remote code execution. This new compromise involves three hidden payload files that run upon Python startup, even if the LiteLLM module isn't explicitly imported. The malicious wheels were found across all major platforms, including macOS and Windows, while the legitimate source distribution lacked these harmful files, highlighting a significant security vulnerability in how Python package integrity is verified. This incident is particularly alarming for the AI/ML community, marking the second known hijacking of LiteLLM in 2026, following a previous compromise that exploited CI/CD pipelines. The malicious code targets sensitive credentials across a wide array of applications and environments, including cloud services and developer tools, making it critical for organizations to reconsider how they trust software distribution. This emphasizes the need for stringent verification processes at the artifact level, rather than relying solely on source code analysis. The presence of executable payloads in wheels underscores the increasing sophistication of supply chain attacks, demanding heightened vigilance and proactive security measures in open-source environments.