LiteLLM hit by second supply chain poisoning attack this year: Steals API keys even without being called

Beating AI Express: Security firm JFrog has discovered that popular open-source AI model gateway LiteLLM has fallen victim to another supply chain attack. Its latest v1.107.0 version was implanted with malware, and official Docker images were also compromised. LiteLLM enables unified access to models from providers like OpenAI and Anthropic, and centralizes API key management. It was previously targeted by a supply chain poisoning attack back in March this year. This latest attack is far more stealthy: hackers hid three extra files in Python packages. After installation, whenever users launch the associated Python environment—even without invoking LiteLLM—the malicious code runs automatically. It steals browser passwords, SSH private keys, AWS and other cloud platform credentials, plus authentication data for development tools including Claude, Codex, and Cursor. The malware also attempts to exfiltrate model API keys stored in LiteLLM’s database to servers masquerading as official domains. Attackers also embedded a persistent backdoor that checks for remote commands hourly, downloading and executing new code as needed. Notably, the official source code on GitHub contained no malicious files, but all seven platform-specific packages released to PyPI were infected with trojans. The community traced the issue via Docker build logs: the release system used a source address pointing to a fake repository that differed from the official one by only one hyphen. How the fake repo slipped into the release pipeline remains unconfirmed. PyPI has since removed the malicious versions, and Docker Hub has taken down affected images. JFrog warns users who installed or ran v1.107.0 to treat all related credentials as potentially compromised, isolate and rebuild their environments, and rotate sensitive data like API keys—rather than simply uninstalling the software. The confirmed safe rollback version is v1.106.2.