LiteLLM compromise exposes developer credentials

Microsoft Threat Intelligence identified a supply chain compromise affecting LiteLLM version 1.107.0, an open-source Python library that connects applications to model APIs from providers including OpenAI, Anthropic and Google. The compromised package used Python startup hooks to silently steal credentials from developer machines and load additional malicious code from a remote server at runtime.

Startup hooks execute automatically when a Python environment initialises, making the malicious behaviour difficult to detect through standard code review. The technique can persist across virtual environments and CI/CD pipelines if not actively detected. Exposed credentials could include cloud service API keys, AI provider accounts, database connections and internal tooling credentials, potentially enabling access to production systems or sensitive data.

Development and security teams should check development, staging and production environments for LiteLLM 1.107.0 and replace affected installations with a clean version. Recommended safeguards include software composition analysis to flag compromised dependencies, secrets management systems instead of environment variables or configuration files, regular dependency audits, pinned versions and integrity verification. AI libraries warrant the same supply chain scrutiny as other open-source dependencies.