Security researchers have disclosed the full scope of the LiteLLM supply-chain poisoning incident that occurred in March, describing it as one of the largest AI-related credential leaks to date. Within roughly 40 minutes, attackers exfiltrated approximately 195TB of credentials belonging to around 2500 organizations, including major technology companies such as Nvidia.
LiteLLM serves as a widely used open-source API gateway for connecting Python applications to services from OpenAI, Anthropic, and Azure. The project records nearly 100 million monthly downloads and is installed approximately 3.4 million times per day, making it a critical piece of AI infrastructure.
Attackers from the group TeamPCP did not target the LiteLLM repository directly. Instead, they compromised a dependency used inside LiteLLM’s CI/CD pipeline: the open-source vulnerability scanner Trivy. By rewriting a Git tag, they injected malicious code into Trivy version v0.69.4. Because LiteLLM’s pipeline pulled the latest Trivy tag without version pinning, the poisoned scanner executed during the build process and stole PyPI publishing tokens.
Using the stolen tokens, the attackers published two malicious LiteLLM releases—1.82.7 and 1.82.8—on PyPI. These packages were downloaded more than 119,000 times before PyPI isolated them. The malware collected and transmitted cloud provider keys, Git tokens, SSH private keys, Kubernetes configurations, environment variables, and API keys for AI services.
Analysis by CloudSEK and Hudson Rock identified roughly 434,000 CI/CD-related credentials within the 195TB dataset. The first user report to PyPI arrived only after one hour and nineteen minutes, underscoring the lack of automated detection for malicious package behavior. A coding flaw in the malware’s .pth launcher triggered recursive fork bombs that crashed victim machines, inadvertently limiting the duration of silent operation.
Security analysts emphasize that modern software supply chains rely on transitive trust across hundreds or thousands of dependencies. In this case, the attack surface extended beyond LiteLLM’s own code to every third-party tool referenced in its build pipeline.
Recommended mitigations include strict version pinning with exact hashes in requirements files, locking GitHub Actions to specific commit hashes rather than moving tags, immediate rotation of any credentials used between March 24 and the discovery date, and deployment of environment-variable exfiltration monitoring inside CI/CD workflows.
The incident demonstrates that AI tooling and infrastructure components are now attractive targets for supply-chain adversaries. Organizations are advised to treat AI dependencies with the same rigor applied to production code and to prepare for an increase in similar attacks against model-serving frameworks, vector databases, and training pipelines.
